Assessment of Pollutant Bioaccumulation in Fish Communities from Human-Induced Water Pollution: A Machine Learning and Statistical Approach

Aquatic animals, particularly fish communities, are highly susceptible to toxic chemicals and the build up of heavy metals resulting from various anthropogenic activities. Despite several individual studies on human and aquatic life, there is a notable gap in integrating a machine learning (ML) approach to assess toxic accumulation differences in different fish sample types, sexes, and ages. Therefore, the objectives of this research are to incorporate ML to perform statistical analyses and determine if there is any dissimilarity in bioaccumulation between different fish sexes, ages, and among different fish sample types. For this purpose, a bid dataset containing 37 variables and 28,616 observations from the Michigan Department of Environment, Great Lakes, and Energy (EGLE, 2024) was used for ML analyses. The dataset included monitoring data on chemicals that bioaccumulate in fish from Michigan waters. Specifically, two-tailed hypothesis tests, neural network analysis, and principal component analysis were conducted to address the objective. The results suggested that older fish accumulated 25.51 times higher levels of DDT than younger fish, and the highest content of toxaphene (0.131 ppb) was observed in fishes with a longer length (69.3 cm) and higher weight (3,225 gm). Hypothesis test results indicated that female fish accumulated significantly higher amounts (0.3±0.19 ppb) of toxic substances, especially mercury, compared to male fish. A neural network analysis with a three-layer network and four nodes produced the best results based on the error percentage. The biplot identified a reciprocal relationship between fish length, weight, and Perfluorooctane sulfonic acid (PFOS) content, while Polychlorinated Biphenyl Congeners (PCBCs) increased with fish length and weight. Overall, this study suggests that hypothesis testing might be accurate for a small subset of data; however, to include the entire dataset, principal component analysis might be a better option depending on the focus of the research.