Wave-driven resuspension as the dominant environmental predictor of enterococci at Chicago beaches.

Predicting fecal indicator bacteria (FIB) to protect public health at Chicago beaches remains a significant challenge due to complex, site-specific environmental dynamics. This study presents a machine learning pipeline utilizing site-specific random forest regressor models to predict log-transformed enterococci concentrations. The models integrate meteorological, hydrodynamic, and temporal-lag features. A hybrid evaluation framework is implemented; models are trained for regression but evaluated on their ability to classify bacterial exceedances at a 320 CCE/100 mL threshold, chosen as a conservative value given data sparsity at the Chicago Park District's Beach Action Value of 1000 CCE/100 mL, which is used to issue swim advisories at these recreational beaches. This is achieved by dynamically identifying an optimal F1 score threshold for each model's raw count predictions, linking regression performance to classification utility. Model interpretability was a primary focus, employing SHAP (SHapley Additive exPlanations) for each beach and in aggregate. This analysis provided critical insights into the distinct drivers at each location. For the Chicago beaches, the models identified hydrodynamic factors, specifically maximum wave height and maximum dominant wave period, as the most significant hydrometeorological predictors of enterococci concentrations, highlighting the importance of nearshore physical processes in this region. The resulting pipeline provides an interpretable, validated, and site-specific forecasting system capable of identifying high-risk contamination events with moderate accuracy but struggling with sensitivity.