Spatial and\nTemporal Analysis, and Machine Learning-Based\nPrediction of PCB Water Concentrations in U.S. Natural Water Systems

Data on dissolved phase water concentrations of polychlorinated\nbiphenyls (PCBs) from 32 locations across the U.S. were compiled from\nreports, Web sites, and peer-reviewed papers, spanning 1979–2020,\nresulting in 5132 individual samples. Data wrangling enabled the organization\nand analysis of this extensive data set. Most samples originated from\nPCB Superfund sites like the Fox, Hudson, and Kalamazoo rivers, New\nBedford Harbor, and Lake Michigan. ΣPCB concentrations ranged\nfrom 10°–10<sup>7.3</sup> pg/L, while individual congener\nmedians ranged from nondetected to 380 pg/L. Non-Aroclor congeners,\ne.g., PCBs 11, 67, and 68, were also reported. Using a machine learning\ntechnique, a Random Forest model accurately predicted the temporal\nand spatial occurrence of dissolved PCBs, achieving Pearson correlations\ngreater than 0.87 for the Anacostia, Fox, Hudson, Kalamazoo, Passaic,\nand Spokane rivers, Chesapeake Bay, and New Bedford Harbor. These\nmodels can be used to forecast PCB concentrations. Through a linear\nmixed-effects model, half-lives of approximately 8 years for ΣPCB\nand individual congeners were determined, but the resulting half-lives\nshowed considerable variability. An interactive map of ΣPCB\nwas created. This investigation highlights the need for additional\nsampling in PCB-contaminated sites that may expose communities to\nairborne PCBs, and in other locations, to enhance our understanding\nof PCB occurrence and distribution.