Development of a Biology-Informed Chemical Mixture Index for Oxidative Stress and Mortality in NHANES 2005-2010: A Survey-Weighted Quantile G-Computation Approach

Background: Current chemical mixture approaches are largely data-driven without considering shared biological mechanisms among mixture components, highlighting the need for biology-informed approaches. Objectives: We constructed an integrated measure of a chemical mixture's oxidative stress potential and assessed its association with mortality in the US population. Methods: The sample comprised 4,574 adults ([≥] 20 years) from National Health and Nutrition Examination Survey (NHANES) 2005-2010. To obtain robust estimates, we performed 1,000 repeated random 50:50 splits into training and testing sets. In each training set, we used survey-weighted quantile g-computation to model serum gamma-glutamyl transferase (GGT), an oxidative stress biomarker, as a function of a 30-chemical mixture (blood metals, urinary polycyclic aromatic hydrocarbons (PAHs), pesticides, phenols/parabens, and phthalates), adjusting for sociodemographic, behavioral, and dietary factors. We then applied the fitted model from each training set to the corresponding testing set to derive the environmental risk score for oxidative stress (ERSOS), defined by predicted GGT values. Associations of ERSOS with all-cause, cardiovascular, and cancer mortality over 11 years of follow-up were estimated in the testing sets using survey-weighted Cox proportional hazards models and summarized across the 1,000 repeated splits. Results: Chemicals with the largest positive weights in quantile g-computation included mono-(2-ethyl-5-hydroxyhexyl) phthalate, mono-2-ethyl-5-carboxypentyl phthalate, 2-hydroxyfluorene, methyl paraben, and benzophenone-3; chemicals with the largest negative weights included mono-(2-ethyl-5-oxohexyl) phthalate and PAH metabolites (1-hydroxynaphthalene, 3-hydroxyphenanthrene, and 3-hydroxyfluorene). The median correlation between observed and predicted GGT in the testing sets was 0.43 (2.5th, 97.5th percentiles: 0.40-0.48). A one standard deviation increase in ERSOS was associated with a median hazard ratio of 1.60 (2.5th, 97.5th percentiles: 1.01-2.57) for cardiovascular mortality. No associations were found for all-cause mortality or cancer mortality. Discussion: The proposed survey-weighted quantile g-computation approach may help estimate biology-informed chemical mixture effects in complex survey data, supporting the potential utility for population-generalizable environmental mixture research.