Machine learning and explainable AI for predicting antibiotics removal in constructed Wetlands: Key factors and management implications

Abstract

Antibiotics are increasingly detected in aquatic environments due to continuous inputs from human and veterinary sources. Constructed wetlands (CWs) are a low-cost and sustainable treatment option for antibiotics, but developing CWs for maximizing their efficiency is challenging due to causative factors. This study applied machine learning (ML) models and explainable artificial intelligence (XAI) to predict removal efficiency of antibiotics in CWs and identify key factors. A dataset of 199 observational cases was compiled from previous literature. Seven factors, including CW type, plant species, hydraulic retention time (HRT), hydraulic loading rate, surface area, influent antibiotic concentration, and antibiotic class, are considered as input variables. Six ML models were applied, and Shapley additive explanations, used as XAI, were applied to identify key causative factors. Among the ML models, CatBoost showed the highest prediction accuracy on the test set (R2 = 0.81). Overall, antibiotic class, influent antibiotic concentration, and HRT were identified as the most influential features, followed by plant species and CW type. This finding suggested that the three controllable variables (CW type, HRT, and plant species) should be carefully considered in CW design to efficiently remove antibiotics. This study elucidated complex relationships between causative factors and CW removal efficiency using MLs, identifying data-driven importance patterns and key interactions influencing removal behavior. The results provide practical ML-based insights for CW design and highlight the potential of ML in managing environmental pollutants with complex behaviors.