A Variance-Based Sensitivity Analysis Approach for Identifying Interactive Exposures.

Area of Science:

• Environmental Health
• Toxicology
• Biostatistics

Background:

• Chemical mixtures pose significant health risks, but analyzing interactions between multiple exposures and health outcomes is complex.
• Bayesian kernel machine regression (BKMR) models nonlinear exposure-health relationships but lacks tools for quantifying interactions.
• Identifying key interacting environmental factors is crucial for public health risk assessment.

Purpose of the Study:

• To develop and validate novel methods for quantifying interactions within Bayesian kernel machine regression (BKMR) models.
• To enable the discovery of high-order interaction terms and rank variable importance in complex exposure scenarios.
• To apply these methods to investigate the interactive effects of environmental exposures on thyroid function in pregnant women.

Main Methods:

• Leveraged the link between BKMR and Gaussian process regressions.
• Adapted variance-based sensitivity analysis tools from uncertainty quantification.
• Proposed a variable clustering approach for interaction quantification and variable ranking.

Main Results:

• Demonstrated method performance through simulations.
• Successfully applied the approach to a real-world dataset.
• Identified interactive effects of per- and polyfluoroalkyl substances (PFAS), diet, and gestational diabetes on thyroid function.

Conclusions:

• The proposed method effectively quantifies interactions and ranks variable importance in BKMR models.
• This approach enhances the understanding of complex chemical mixture effects on human health.
• Provides a valuable tool for environmental health research, particularly in prenatal exposure studies.