Large-scale assessment of PFAS compounds in drinking water sources using machine learning

Area of Science:

• Environmental Chemistry
• Environmental Science
• Public Health

Background:

• Per- and Polyfluoroalkyl substances (PFAS) are a growing public health concern in drinking water.
• Existing methods for assessing PFAS presence and concentration have limitations in reliability and applicability across different regions.
• Understanding the key variables influencing PFAS distribution is crucial for effective water source management.

Purpose of the Study:

• To evaluate the validity of existing PFAS predictive models and explanatory variables in a new geographical area.
• To develop a novel, more reliable predictive model for regional PFAS occurrence and concentration.
• To improve the accuracy and efficiency of PFAS monitoring in drinking water.

Main Methods:

• Reconstruction of four advanced models (spatial regression, random forest, boosted regression trees) using Michigan's statewide dataset.
• Inclusion of diverse explanatory variables: local soil, hydrology, and proximity to contamination sources.
• Application of Bayesian variable selection and a hybrid machine learning-CAR model for PFAS assessment.

Main Results:

• PFAS occurrence was predicted with >90% accuracy, comparable to existing models but using fewer variables.
• Estimated PFAS concentrations showed high alignment with observations (ρ > 0.90, R² > 0.77), outperforming previous methods.
• The novel model demonstrated effectiveness with low data requirements.

Conclusions:

• The developed hybrid model offers a reliable and efficient approach for regional PFAS prediction.
• The findings support improved strategies for monitoring and managing PFAS contamination in drinking water sources.
• This research contributes to a better understanding of PFAS behavior and distribution at a regional scale.