Scholarlink

Background <p>Arsenic contamination in freshwater systems of low- and middle-income regions remains a persistent public health concern, particularly in mining-impacted basins where hydrogeochemical processes, seasonal variability, and anthropogenic pressures interact in complex, nonlinear ways. Existing approaches often separate spatial modelling, prediction, and attribution, limiting integrated interpretation. This study develops a unified, uncertainty-aware framework to predict and characterize arsenic risk in Ghana’s Pra–Ankobra–Tano river basins, while providing mechanistically consistent, but not confirmatory, insights into potential drivers.</p> Methods <p>A multi-component analytical system was implemented using 100 in-situ observations (January–April 2020). Spatiotemporal variability was modelled using a Gaussian-process generalized additive model (GP–GAM), while extreme events (≥ 90th percentile; ~ 19.54&#xa0;µg/L) were predicted via tuned XGBoost with SHAP-based interpretability. Generalized random forests were applied to estimate temperature-associated contrasts under an observational framework. Gaussian mixture models identified geochemical regimes, ROC analysis optimized thresholds, and spatial diagnostics (Moran’s I, Getis–Ord Gi*) and simulation-based surfaces were used to characterize spatial structure and uncertainty.</p> Results <p>Predicted arsenic concentrations exhibited spatial gradients (~ 9–13&#xa0;µg/L, within the masked domain, interpreted as data-supported patterns rather than fixed structures, with elevated exceedance probabilities concentrated in central–southeastern zones. The XGBoost model demonstrated moderate discrimination under cross-validation (AUC up to ~ 0.83 with variability), with temperature and pH emerging as dominant predictors of spike risk. Estimated temperature-associated contrasts were positive (ATE ≈ 5.70&#xa0;µg/L; 95% CI 3.64–7.76), indicating higher concentrations under warmer conditions within the modelled framework. Two geochemical regimes were identified, distinguishing lower- and higher-concentration profiles. Counterfactual simulations (+ 2&#xa0;°C, − 0.5 pH) suggested modest increases in predicted arsenic levels (~ 0.90–1.18&#xa0;µg/L), with spatial heterogeneity.</p> Conclusion <p>The integrated framework provides an integrated and uncertainty-aware analytical system for arsenic risk assessment, enabling simultaneous prediction, attribution, and spatial diagnostics. Findings are consistent with temperature-sensitive arsenic dynamics but should be interpreted as mechanistically plausible patterns derived from observational data rather than definitive causal or process-based confirmation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A unified geostatistical machine learning framework for predicting and attributing arsenic contamination in southwestern Ghana

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月21日

查看全文
  • Senyefia Bosson-Amedenu,
  • Emmanuel Ayitey,
  • Bridget Sena Borbor

摘要

Background

Arsenic contamination in freshwater systems of low- and middle-income regions remains a persistent public health concern, particularly in mining-impacted basins where hydrogeochemical processes, seasonal variability, and anthropogenic pressures interact in complex, nonlinear ways. Existing approaches often separate spatial modelling, prediction, and attribution, limiting integrated interpretation. This study develops a unified, uncertainty-aware framework to predict and characterize arsenic risk in Ghana’s Pra–Ankobra–Tano river basins, while providing mechanistically consistent, but not confirmatory, insights into potential drivers.

Methods

A multi-component analytical system was implemented using 100 in-situ observations (January–April 2020). Spatiotemporal variability was modelled using a Gaussian-process generalized additive model (GP–GAM), while extreme events (≥ 90th percentile; ~ 19.54 µg/L) were predicted via tuned XGBoost with SHAP-based interpretability. Generalized random forests were applied to estimate temperature-associated contrasts under an observational framework. Gaussian mixture models identified geochemical regimes, ROC analysis optimized thresholds, and spatial diagnostics (Moran’s I, Getis–Ord Gi*) and simulation-based surfaces were used to characterize spatial structure and uncertainty.

Results

Predicted arsenic concentrations exhibited spatial gradients (~ 9–13 µg/L, within the masked domain, interpreted as data-supported patterns rather than fixed structures, with elevated exceedance probabilities concentrated in central–southeastern zones. The XGBoost model demonstrated moderate discrimination under cross-validation (AUC up to ~ 0.83 with variability), with temperature and pH emerging as dominant predictors of spike risk. Estimated temperature-associated contrasts were positive (ATE ≈ 5.70 µg/L; 95% CI 3.64–7.76), indicating higher concentrations under warmer conditions within the modelled framework. Two geochemical regimes were identified, distinguishing lower- and higher-concentration profiles. Counterfactual simulations (+ 2 °C, − 0.5 pH) suggested modest increases in predicted arsenic levels (~ 0.90–1.18 µg/L), with spatial heterogeneity.

Conclusion

The integrated framework provides an integrated and uncertainty-aware analytical system for arsenic risk assessment, enabling simultaneous prediction, attribution, and spatial diagnostics. Findings are consistent with temperature-sensitive arsenic dynamics but should be interpreted as mechanistically plausible patterns derived from observational data rather than definitive causal or process-based confirmation.