Context-dependent removal of emerging pollutants in constructed wetlands: a dependence-aware meta-analysis and explainable machine-learning framework for performance envelopes.

Constructed wetlands (CWs) are widely used as low-energy polishing systems, yet their performance for emerging pollutants (EPs) varies across pollutant classes, wetland configurations, wastewater matrices, and experimental scales. We synthesized 67 peer-reviewed studies and 849 paired influent-effluent comparisons using multilevel random-effects models with study-level cluster-robust inference. Heterogeneity was extreme (I2≈99%), and pollutant class, flow configuration, wastewater matrix, and scale defined distinct attenuation regimes. Microplastics showed the strongest short-term outlet attenuation, primarily reflecting physical retention within wetland compartments under the available monitoring durations. None of the 14 continuous moderators remained significant after Benjamini-Hochberg false-discovery-rate correction, indicating that performance is shaped by combinations of conditions rather than a single universal control variable. In the study-grouped held-out evaluation, CatBoost outperformed the linear and other baseline models (R2 = 0.619; RMSE = 1.251). Leave-one-study-out validation retained a positive but lower predictive signal (R2 = 0.238; RMSE = 3.125; MAE = 2.055). We translate these findings into quantitative, context-specific performance envelopes that report expected lnRR together with confidence and prediction intervals. The resulting framework supports scenario-based system selection, conservative scale-up, and targeted monitoring of hydraulic conditions, wastewater chemistry, and particulate retention.