2598

Environmental health researchers may aim to identify sources (i.e., patterns), such as product use or behaviors, that give rise to mixtures of potentially harmful environmental exposures such as endocrine disrupting chemicals (EDCs). Existing methods are limited by user-specified number and interpretability of patterns in terms of human understanding.
We adapted a non-parametric Bayesian Non-negative Matrix Factorization (npBNMF) to identify patterns of chemical exposures when the number of patterns is not known a priori. We placed strictly positive continuous priors on pattern loadings and individual scores, making both interpretable as concentrations, and used a non-parametric prior to estimate the pattern number. To validate this method, we simulated 300 datasets with increasing levels of complexity: (1) distinct underlying patterns, (2) overlapping patterns, and (3) overlapping patterns and correlated scores. We compared npBNMF performance with non-negative matrix factorization (NMF), principal component analysis (PCA), and factor analysis (FA). After validation, we applied npBNMF to 51 EDCs in 569 pregnant mothers in the Columbia Center for Children’s Environmental Health “Mothers and Newborns Cohort.”
In simulations, npBNMF out-performed traditional methods and estimated the true number of patterns in 98% of runs, something no other known method can do. npBNMF and NMF provided non-negative scores and loadings with no orthogonality constraint, thus providing more interpretable results than PCA, whose mutually uncorrelated patterns are implausible in environmental health, or FA, whose negative scores and loadings are incomprehensible as concentrations.
We identified three patterns of EDC exposure in pregnant mothers, corresponding well with known EDC classes. Additionally, we observed pattern associations with measured behaviors, such as personal care product use.
npBNMF successfully identified EDC patterns that correspond well with current understanding of EDC sources. In future research, we will examine associations between patterns and health outcomes to inform interventions. npBNMF may be used to identify patterns in environmental mixtures.