Predicting Groundwater PFOA Exposure Risks with Bayesian Networks: Empirical Impact of Data Preprocessing on Model...

Area of Science:

• Environmental Science
• Data Science
• Toxicology

Background:

• Per- and polyfluoroalkyl substances (PFAS) are widespread environmental contaminants of growing concern.
• Machine learning models can predict PFAS exposure risks using environmental data.
• Data preprocessing methods can influence model outcomes, but their effects are not well understood.

Purpose of the Study:

• To evaluate how different data preprocessing techniques affect machine-learned Bayesian network models for predicting perfluorooctanoic acid (PFOA) in groundwater.
• To assess the impact of data preprocessing on model accuracy and the identification of high-risk exposure areas.

Main Methods:

• Utilized 19 years of PFOA measurements from Minnesota, USA, linked to potential sources and environmental fate factors.
• Applied nine distinct data preprocessing methods to create training datasets.
• Trained Bayesian network models to predict the probability of PFOA concentrations exceeding the Minnesota health advisory level (35 ppt).

Main Results:

• Varying preprocessing approaches resulted in different model structures and accuracies.
• All models demonstrated robust performance, accurately distinguishing between high-risk and low-risk wells (82.0%–89.0% accuracy).
• A trade-off exists between data quality and model performance; stricter screening reduced sample size but potentially improved data quality.

Conclusions:

• Data preprocessing is a critical step that influences the structure and accuracy of machine learning models for PFAS exposure assessment.
• Despite preprocessing variations, the developed models effectively identified groundwater wells at high risk for PFOA contamination.
• Optimizing data preprocessing strategies is essential for reliable PFAS risk prediction, balancing data quantity with quality.