How to assess exposure to atmospheric pollutants in epidemiological studies? An illustration of some trade-offs in the context of pregnancy exposures

摘要

Background: The concerns regarding possible effects of maternal exposure to atmospheric pollutants on pregnancy outcomes warrant effort in exposure assessment. Aims: We review and discuss the advantages and limitations of the various tools considered in epidemiology to assess air pollution exposure during pregnancy, focusing on outdoor models and tools for personalised exposure assessment. Methods and results: Models based on air quality monitoring station (AQMS) networks have a poor spatial resolution. Land-use regression (LUR) models have an improved spatial resolution but classically provide a yearly exposure estimate. We have suggested to combine LUR models with AQMS data to increase their temporal resolution. As shown in the context of Eden mother-child cohort, these seasonally-adjusted LUR models are in good agreement with dispersion models to predict pregnancy average exposure to pollutants such as NO2. The main limitations of such approaches are that 1) space-time activity and 2) indoor exposures are not taken into account. In a newly set mother-child cohort, SEPAGES-feasibility, women carried GPS and personal NO2 dosimeters and indoor measures have been conducted. We provide estimates of the amount of exposure misclassification attributable to these two sources of error, compared to approaches based solely on outdoors model-based estimates at the home address. Discussion: The progress in the assessment of exposure of pregnant women to atmospheric pollutants in the last decades have allowed refining exposure estimates and integrating time-, space-and behaviour-related exposure contrasts, at the cost of increase in logistic and budget constraints. These approaches may be cumbersome to implement on a large scale (to study rare outcomes); for such outcomes, detailed assessment of exposure using personal tools in a sub-population, combined to measurement error models to translate results in the larger population may be a good compromise.