Predicting Arsenic Bioaccessibility: A Global Data-Driven Machine Learning Approach and Its Implication for Reducing Carbon Emissions

Abstract

Site-specific arsenic (As) bioaccessibility data can improve the accuracy of health risk assessments, but direct measurements are costly and time-consuming. Even when available, measured values such as the mean still yield remediation targets below natural background levels, limiting their practical use. Existing predictive models, including linear regressions and some machine learning (ML) approaches, often rely on artificially spiked or limited field-aged samples with high As concentrations, reducing their generalizability. A global dataset of 1,458 records of As bioaccessibility in field-aged soils from studies since the 1990s were complied, covering a wide range of As concentrations (As-T) and soil properties. Gastric bioaccessibility showed a log-normal distribution with a mean of 23.4%. Among eight ML models, the Random Forest (RF) model performed best (R² = 0.86, RMSE = 0.58). As-T explained 73.2% of the variance, with significant relationships observed with Fe, Mn, organic carbon, and pH. Applied to a contaminated sintering site in southwest China, the RF-informed probabilistic risk assessment yielded a remediation target three times higher than current standards and reduced soil remediation volume and carbon emissions by 79.1%. This study highlights the potential of ML to enhance risk assessment accuracy and support more sustainable site remediation strategies.