Abstract
Per- and polyfluoroalkyl substances (PFAS) are contaminants that can lead to adverse health effects in aquatic organisms, including reproductive toxicity and developmental abnormalities. To assess the ecological health risk of PFAS in Pennsylvania stream surface water, we conducted a comprehensive analysis that included both measured and predicted estimates. The potential combined exposure effects of 14 individual PFAS to aquatic biota were estimated using the sum of exposure-activity ratios (ΣEARs) in 280 streams. Additionally, machine learning techniques were utilized to predict potential PFAS exposure effects in unmonitored stream reaches, considering factors such as land use, climate, and geology. Leveraging a tailored convolutional neural network (CNN), a validation accuracy of 78% was achieved, directly outperforming traditional methods that were also used, such as logistic regression and gradient boosting (accuracies of ~65%). Feature importance analysis highlighted key variables that contributed to the CNN's predictive power. The most influential features highlighted the complex interplay of anthropogenic and environmental factors contributing to PFAS contamination in surface waters. Industrial and urban land cover, rainfall intensity, underlying geology, agricultural factors, and their interactions emerged as key determinants. These findings may help to inform biotic sampling strategies, water quality monitoring efforts, and policy decisions aimed to mitigate the ecological impacts of PFAS in surface waters.