Characterizing multipollutant exposure profiles among a cohort of pregnant women via self-organizing maps

Abstract

Exposure to pollutants during pregnancy can adversely impact fetal development during the sensitive prenatal period. Such exposures do not occur in isolation, and pregnant women are exposed to multipollutant mixtures, which poses challenges for the investigation of risks of harm of environmental pollutants. One approach to investigate health effects of multipollutant mixtures is to derive exposure profiles from multipollutant exposure data, which obviates the challenging parameterization of multipollutant exposures and enable contrasts of health outcomes across exposure profiles. A promising method for exposure profile discovery and characterization is the self-organizing map (SOM). SOMs are similar to the k-means clustering algorithm, with the added utility of a spatially correlated topology among clusters that enhances understanding of between-cluster relationships. SOMs were used to characterize multipollutant exposure profiles among pregnant women from the New Hampshire Birth Cohort Study (n=337). Multipollutant exposures were assessed via silicone wristband passive monitors, worn for 7±1 days at 15±6 gestational weeks. Wristbands were analyzed for concentrations of 1530 organic pollutants via gas chromatography. Of 200 chemicals detected in at least one wristband, 18 were detected in ≥60% of wristbands. These 18 chemicals included five phthalates, six pesticides, and several chemicals used in consumer products and personal care products. A rectangular SOM of dimensions 7x2 was fit to these frequently detected chemicals. The 14 SOM profiles reflected unique combinations of these chemicals and ranged in size from 180 to 1 participants per profile. The largest profile was characterized by low to moderate exposures to most pollutants, whereas other profiles had exceptionally high or low exposure to a subset of pollutants. Certain covariates varied across SOM profiles, including BMI, educational attainment, race/ethnicity, and season of wear. This work demonstrates the utility of SOMs for discovering and characterizing exposure profiles from multifaceted exposure data.