Principal Component Pursuit for Pattern Identification in Environmental Mixtures.

Area of Science:

• Environmental health
• Computational toxicology
• Statistical modeling

Background:

• Identifying sources of environmental exposures is crucial for public health.
• Environmental mixtures often contain complex patterns of co-occurring pollutants.

Purpose of the Study:

• To adapt and evaluate principal component pursuit (PCP) for analyzing environmental exposure mixtures.
• To compare the performance of PCP with limit of detection (PCP-LOD) against principal component analysis (PCA).

Main Methods:

• Adapted PCP to handle non-negative, missing, and below-limit-of-detection data.
• Simulated environmental mixture data with varying proportions and noise structures.
• Applied PCP-LOD to a National Health and Nutrition Examination Survey (NHANES) dataset of persistent organic pollutants (POPs) in U.S. adults.

Main Results:

• PCP-LOD accurately recovered the true number of exposure patterns in simulations, unlike PCA.
• PCP-LOD demonstrated lower predictive error than PCA, especially with up to 50% missing data.
• Analysis of POPs revealed a rank-three structure, with patterns reflecting comprehensive, structure-based, and toxicity-based exposures.

Conclusions:

• PCP-LOD is a valuable tool for characterizing complex environmental exposures.
• Identified exposure patterns can inform targeted public health interventions and messaging.