Endocrine disrupting chemicals (EDCs) include multiple classes of chemicals that have been used extensively in consumer products. Mounting evidence from toxicological and epidemiological studies suggest EDCs are developmental neurotoxicants, and EDC exposure during the critical in utero period is associated with adverse child cognitive development. Unfortunately, current research focuses on individual EDCs and largely ignores joint and interactive effects of EDCs and the overall effect of the EDC mixture. To assess exposure to multiple EDCs simultaneously, one must consider the high dimensionality of the exposure matrix and the complex correlation structures across chemicals in statistical analyses. To address limitations of existing methods, we propose to adapt a robust technique that is well-established for pattern recognition and dimensionality reduction in machine learning. We specifically aim to use Latent Dirichlet Allocation (LDA), a type of robust Bayesian non-negative matrix factorization, to determine the patterns of exposure to four ubiquitous classes of EDCs known to cross the placenta?polybrominated diphenyl ethers (PBDEs), polychlorinated biphenyls (PCBs), phenols (e.g., bisphenol A), and phthalates?and to characterize the relationship between these exposure patterns and cognitive development. LDA is empirically-driven so that the researcher does not need to specify a priori the number of patterns, and the non-negativity constraint enhances the interpretability of the patterns identified. For our health model, we will use a supervised approach that allows child cognitive scores to inform the LDA solution, thereby enabling identification of patterns most relevant to the outcome. We will conduct this work using the existing infrastructure of the Columbia Center for Children?s Environmental Health Mothers and Newborns Study, a longitudinal birth cohort of mother-child dyads. We will also establish reproducibility of the method by creating a user-friendly R package so that other researchers can easily apply LDA in environmental epidemiology, and we will verify transferability and functionality of the method on a separate cohort. This will be the first study to assess the interacting and overall effects of multiple EDCs on child cognitive development, introducing LDA as a straight-forward tool for the analysis of complex mixtures in epidemiology. If successful, this method has broader implications for environmental epidemiology, as it can easily be applied to other environmental mixtures of interest. The activities encompassed by this proposal (study design, data management, advanced statistics, machine learning, data science, and presentation of findings) cover the set of fundamental research skills required by a scientist entering the interdisciplinary field of environmental epidemiology in the era of Big Data and Precision Public Health. The applied experience gained from carrying out this research, in combination with didactic training and individual cross-disciplinary mentoring, comprises a comprehensive research training plan that will serve as a platform from which to launch a career as an independent investigator in environmental epidemiology.
This study will be the first to determine existing patterns of cross-class endocrine disrupting chemical (EDC) exposure in utero and to evaluate their association with child cognitive development, assessing potentially non- linear independent effects, non-additive interactions, and the overall effect of the mixture. This will help to identify environmental risk factors most amenable to public health interventions and inform future researchers and policy makers about concurrent exposure to EDCs in a vulnerable population The adaptation of Latent Dirichlet Allocation (LDA) for pattern identification and dimension reduction in environmental epidemiology integrates data science and machine learning into environmental health research, which is novel and extremely relevant to public health, as LDA can be easily applied to other environmental mixtures and co-exposures of interest.
