Prediction of 35 Target Per- and Polyfluoroalkyl Substances (PFASs) in California Groundwater Using Multilabel Semisupervised Machine Learning

Abstract

Comprehensive monitoring of perfluoroalkyl and polyfluoroalkyl substances (PFASs) is challenging because of the high analytical cost and an increasing number of analytes. We developed a machine learning pipeline to understand environmental features influencing PFAS profiles in groundwater. By examining 23 public data sets (2016−2022) in California, we built a state-wide groundwater database (25,000 observations across 4200 wells) encompassing contamination sources, weather, air quality, soil, hydrology, and groundwater quality (PFASs and cocontaminants). We used supervised learning to prescreen total PFAS concentrations above 70 ng/L and multilabel semisupervised learning to predict 35 individual PFAS concentrations above 2 ng/L. Random forest with ADASYN oversampling performed the best for total PFASs (AUROC 99%). XGBoost with SMOTE oversampling achieved the AUROC of 73−100% for individual PFAS prediction. Contamination sources and soil variables contributed the most to accuracy. Individual PFASs were strongly correlated within each PFAS's subfamily (i.e., short-vs long-chain PFCAs, sulfonamides). These associations improved prediction performance using classifier chains, which predicts a PFAS based on previously predicted species. We applied the model to reconstruct PFAS profiles in groundwater wells with missing data in previous years. Our approach can complement monitoring programs of environmental agencies to validate previous investigation results and prioritize sites for future PFAS sampling.

Application Of Machine Learning To Understand Pfas Occurrence, Distribution, Transport And Removal In Water

“…The traintest split ratio is the percentage of the data that would be used to train the model to that which would be used to test it. Split such as 80:20 was used by several authors (Azhagiya Singam et al, 2020;Dong et al, 2023;Hosseinzadeh et al, 2022;McMahon et al, 2022). Hu et al, (2021) divided the domestic well PFAS data in the ratio 80:20; this is consistent with the ratio adopted by DeLuca et al, (2023) who used a 100-iteration Monte Carlo holdout scheme to split PFAS data from Columbia River Basin fish tissue.…”

“…This process is repeated several times (iteration number), changing the subset each time. Yuan et al, (2023) used a 5-fold CV, and George and Dixit, (2021) used 10 subsets (10-fold CV) to group groundwater wells to prioritize PFAS testing; the same number was used by Dong et al, (2023) for total PFAS prediction. In addition to this, Cao et al, (2023) performed 500 iterations on the training set for hyperparameter tuning.…”

“…George and Dixit, (2021) used GAMA data to perform a predictive model on prioritizing groundwater testing for all wells in California. Dong et al, (2023) used 12 groundwater data to predict 35 target PFAS in California.…”

“…Kwon et al, (2023) predicted bioactivities of PFAS. Other data sources include the United states environmental protection agency's water quality portal (USEPA) (Azhagiya Singam et al, 2020;DeLuca et al, 2023;Dong et al, 2023), PubChem Bioassay Database (Kwon et al, 2023), Pennsylvania Water quality network (Breitmeyer et al, 2023), from previously published studies on PFAS (Karbassiyazdi et al, 2022;Kibbey et al, 2020;Patel et al, 2022), lake and river data (Antell et al, 2023;Stults et al, 2023), Minnesota department of health (MDH) Government agency data (Breitmeyer et al, 2023;Fernandez et al, 2023;Li and Gibson, 2023) and experimental data (Cao et al, 2022;Sörengård et al, 2022;Wang et al, 2022). Some authors combined several public data for their machine learning predictions.…”

Applications of Artificial Intelligence, Machine Learning, and Data Analytics in Water Environments

“…The special issue includes several review articles encompassing a wide spectrum, ranging from a historical perspective of water data to computational modeling in wastewater treatment to ML modeling of environmental chemical reactions, environmental toxicology, heavy metal removal, and cyanobacterial harmful algal blooms (HABs) . One significant application of these innovative tools is ML-assisted environmental monitoring, which can address diverse problems, such as predicting effluent nutrients or influent flow rates and nutrient loads at wastewater treatment plants, , formation of disinfection byproducts, drivers of the accumulation of potentially toxic elements in sediments, greenhouse gas emissions, , occurrence of PFAS, water quality assessment, microplastics, microcystins, and differentiation of landfill leachate and domestic sludge . ML has also been extensively employed to model environmental chemical reactions and processes, including adsorption onto various materials, , biodegradation, photodegradation, and the physicochemical and meteorological variables that affect the seasonal growth and decline of HABs .…”