Machine learning–driven identification of key factors governing hormone adsorption by agrifood waste–derived adsorbents

Hormone contamination in aquatic systems threatens human health and ecosystems, demanding efficient and sustainable separation strategies. This study aims to develop a machine learning (ML) framework to predict hormone adsorption using agrifood waste-derived adsorbents and to identify the key variables governing adsorption efficiency. A dataset of 604 observations with nine input variables was compiled from peer-reviewed studies. Three ensemble ML algorithms—Random Forest (RF), Least Squares Boosting, and M5 model trees—were evaluated across 511 models generated from all combinations of one to nine input variables. RF demonstrated the highest predictive accuracy, achieving a maximum test coefficient of determination (R 2 ) of 0.976 and a minimum mean absolute error of 4.26 using eight input variables. Model robustness was assessed using a Combined Index (CI) integrating six statistical metrics, which decreased from a median value of 0.78 for single-input models to 0.0006 for the optimal eight-input configuration; the full nine-input model yielded CI = 0.0019. Bayesian optimization using the Expected Improvement per Second acquisition function further improved model performance, producing the highest test Nash–Sutcliffe efficiency (0.951) and the lowest normalized root mean square error (0.162) and ratio of root mean square error to standard deviation (0.221). Feature importance analysis identified initial concentration, contact time, pH, and adsorbent-to-solution ratio as the most influential predictors of adsorption efficiency. Hormone-specific analyses for 17β-estradiol (318 data points) and 17α-ethinylestradiol (106 data points) confirmed the robustness of these findings. The results show that optimized ML models support the design of sustainable adsorption-based hormone removal systems. • ML–Bayesian optimization (BO) framework built from a 604-point literature dataset. • EIPS-based BO maximized NSE (0.951) and minimized prediction errors. • Random Forest showed highest accuracy via optimized hyperparameter selection. • Initial concentration, contact time, pH, and adsorbent ratio govern adsorption. • Consistent key factors confirmed for estradiol and ethinylestradiol subgroups.