Background Exposure to nitrogen dioxide (NO2), ozone (O3), fine particulate matter (PM2.5), and heat has previously been associated with preterm birth. However, individuals are often exposed to a mixture of pollutants, which may exacerbate health effects. Unsupervised neural networks are highly suitable for characterizing high-dimensional exposures, but few studies of environmental mixtures have utilized these algorithms. Objectives To implement a novel epidemiologic machine learning framework for linking high-dimensional mixture exposures and health outcomes, with an application in preterm birth and residential exposure to PM2.5, O3, NO2, and temperature. Methods Data comes from a retrospective cohort of 44,874 individuals living in Utah who gave birth for the first time between 2013 and 2016. Fine-scale air pollution estimates were used to create a high-dimensional self-organizing map (a type of unsupervised neural network) of exposure mixtures. We used cluster analysis to group similar weekly exposure mixtures. Pregnancies were linked to the clusters, and Bayesian mixed-effects logistic regression was used to estimate odds ratios at each gestational week with false discovery rate (FDR) multiple comparison adjustments. Results Exposure to certain mixtures at critical windows was associated with a higher risk of preterm birth. In particular, maternal exposure to Cluster 10 (a high O3 and PM2.5 mixture) in weeks 9–14 was associated with up to 53% greater risk of preterm birth, peaking in week 11 (ORweek11: 1.53, 95% BCI [1.12, 2.08], Cr = 99.7%). Repeated exposure for this entire period (weeks 9–14) was associated with 2.8-times greater risk of preterm birth (OR: 2.81, 95% BCI [1.99, 3.96], Cr = 100%). Significance Using a novel machine learning approach, we identified several patterns of exposure to mixtures, primarily composed of O3 and PM2.5, which may be associated with preterm birth at critical windows. The proposed framework reduces complexity while preserving time-varying exposures. Impact <UnorderedList Mark="Bullet"> <ItemContent> Exposure to air pollution mixtures at critical times during pregnancy may have synergistic effects on the risk of preterm birth, and few studies have examined these relationships using neural networks. We propose an epidemiologic machine learning framework for dimension reduction, which assigns exposure types to individuals over time and allows for estimation of odds ratios of the outcome of interest. Using this method, we describe patterns of exposure to mixtures of NO2, O3, PM2.5, and heat and estimate their effects on preterm birth. </ItemContent> </UnorderedList>

Linking mixtures of air pollution exposures and preterm birth with a self-organizing map

期刊论文
发表日期：

2026年6月22日

查看全文

Brenna C. Kelly,
Simon C. Brewer,
Joel D. Schwartz,
Robert M. Silver,
Heather A. Holmes,
Heidi A. Hanson,
Jennifer A. Doherty,
Michelle P. Debbink

摘要

Background

Exposure to nitrogen dioxide (NO₂), ozone (O₃), fine particulate matter (PM_2.5), and heat has previously been associated with preterm birth. However, individuals are often exposed to a mixture of pollutants, which may exacerbate health effects. Unsupervised neural networks are highly suitable for characterizing high-dimensional exposures, but few studies of environmental mixtures have utilized these algorithms.

Objectives

To implement a novel epidemiologic machine learning framework for linking high-dimensional mixture exposures and health outcomes, with an application in preterm birth and residential exposure to PM_2.5, O₃, NO₂, and temperature.

Methods

Data comes from a retrospective cohort of 44,874 individuals living in Utah who gave birth for the first time between 2013 and 2016. Fine-scale air pollution estimates were used to create a high-dimensional self-organizing map (a type of unsupervised neural network) of exposure mixtures. We used cluster analysis to group similar weekly exposure mixtures. Pregnancies were linked to the clusters, and Bayesian mixed-effects logistic regression was used to estimate odds ratios at each gestational week with false discovery rate (FDR) multiple comparison adjustments.

Results

Exposure to certain mixtures at critical windows was associated with a higher risk of preterm birth. In particular, maternal exposure to Cluster 10 (a high O₃ and PM_2.5 mixture) in weeks 9–14 was associated with up to 53% greater risk of preterm birth, peaking in week 11 (OR_week11: 1.53, 95% BCI [1.12, 2.08], Cr = 99.7%). Repeated exposure for this entire period (weeks 9–14) was associated with 2.8-times greater risk of preterm birth (OR: 2.81, 95% BCI [1.99, 3.96], Cr = 100%).

Significance

Using a novel machine learning approach, we identified several patterns of exposure to mixtures, primarily composed of O3 and PM2.5, which may be associated with preterm birth at critical windows. The proposed framework reduces complexity while preserving time-varying exposures.

Impact

Exposure to air pollution mixtures at critical times during pregnancy may have synergistic effects on the risk of preterm birth, and few studies have examined these relationships using neural networks. We propose an epidemiologic machine learning framework for dimension reduction, which assigns exposure types to individuals over time and allows for estimation of odds ratios of the outcome of interest. Using this method, we describe patterns of exposure to mixtures of NO₂, O₃, PM_2.5, and heat and estimate their effects on preterm birth.