Predicting phthalate release from microplastics and screening associated ecological risks using interpretable machine learning

Phthalate esters (PAEs) are widely used as plastic additives and can leach from microplastics (MPs) into aquatic environments, making MPs a potentially overlooked secondary source of PAEs. However, comparisons across studies and risk prioritization remain difficult because published release data were generated under highly heterogeneous experimental conditions. Here, we compiled 701 observations from published studies of PAE leaching from MPs in aqueous systems and developed an interpretable machine-learning framework for screening-level estimation of mass-normalized PAE release capacity (RC) within the represented data domain. Among the five algorithms evaluated, XGBoost achieved the best test-set performance (R² = 0.800, RMSE = 1.032, and MAE = 0.668). SHAP analysis showed that logKow, polymer type, and particle size are the most influential features for predicting ln(RC). Application of the framework to five aquatic scenarios showed that screening-level ecological risk was not explained by total MP abundance alone but varied with polymer-PAE combinations, particularly those involving PVC. Di(2-ethylhexyl) phthalate (DEHP) dominated the estimated risk across all scenarios, accounting for 91-97% of total RQ. These results highlight that total MP abundance alone is insufficient for screening-level risk estimation, providing a data-driven basis for prioritizing potentially important polymer-PAE combinations in aquatic monitoring.