Ecological and health risk assessment of soil metal(loid)s based on machine learning approaches

Ecological risk assessment (ERA) and health risk assessment (HRA) of metal(loid)-contaminated soils are essential for the achievement of risk mitigation and sustainable soil management. While conventional risk assessment methodologies are constrained by static modeling and a lack of multivariate considerations, machine learning (ML) has demonstrated significant potential in soil metal(loid) ERA and HRA by leveraging robust data mining and pattern recognition capabilities. Following the PRISMA guidelines, the review included 54 peer-reviewed studies regarding ML-based ERA and HRA in metal(loid)-contaminated soils from the Web of Science and Scopus databases. The analysis of 54 reviewed studies reveals an upward trend in research within this field from 2021 to 2025, focusing on Cd, Pb, As, Cr, Cu, Zn, Ni, and Hg. Input data for ML approaches include environmental covariates (94.44%) and hyperspectral data (5.56%). ML applications in risk assessment comprise four distinct pathways: risk assessment based on direct modeling of risk indicators (29.23%), risk assessment based on predicted concentrations (32.31%), source-oriented risk assessments (32.31%), and risk assessments improved by environmental factors (6.15%). Specifically, ERA is mainly applied through direct modeling of risk indicators (42.86%), whereas source-oriented risk assessment (46.67%) is the primary pathway for HRA. This review provides a comprehensive synthesis of the current status and developmental trends, an overview and selection guidelines for ML approaches, a summary of common input variables, and a detailed analysis of application pathways, while addressing existing challenges and future perspectives.