Using interpretable machine learning method to predict children’s fluoride exposure health risks based on routine water quality data: An empirical study in the Huaihe River Basin, China.

Introduction

Fluoride is a naturally occurring element widely distributed in aquatic environments and is commonly detected in both surface water and groundwater systems (Liu et al., 2024; Sun et al., 2025; T. Wang et al., 2020). While low levels of fluoride are considered beneficial for dental health, excessive fluoride exposure has been consistently associated with adverse health outcomes, including dental fluorosis, skeletal fluorosis, and potential effects on neurodevelopment (Johnston and Strobel, 2020; Mazzoli et al., 2025). Drinking water represents one of the primary exposure pathways for fluoride in human populations, making fluoride contamination a persistent concern in environmental health research and water resource management.

Children are particularly vulnerable to fluoride exposure due to their higher water intake relative to body weight, rapid skeletal development, and immature physiological detoxification capacity (Liu et al., 2019; M. W. Wang et al., 2020). Numerous epidemiological and toxicological studies have demonstrated that chronic fluoride intake during early life stages can lead to irreversible dental and skeletal damage and may be associated with adverse developmental effects (Alvarez et al., 2009; Ismail and Bandekar, 1999; Zhou et al., 2023). Consequently, health risk assessment frameworks increasingly emphasize children as a priority population when evaluating fluoride exposure through drinking water (Din et al., 2024; Jannat et al., 2022; Tokath et al., 2022; Tokatli et al., 2024). However, translating fluoride occurrence data into actionable health risk information for children remains challenging, especially at large spatial and temporal scales.

Existing research typically assesses static health risks by characterizing fluoride concentration levels in specific areas (Liu et al., 2021; Qasemi et al., 2019; K. Zhang et al., 2020). Mu et al. (2024) evaluated the human health risks of fluoride in urban river areas; Q. Y. Zhang et al. (2020) conducted a similar assessment the human health risks of fluoride in the Jiaokou Irrigation District of China; Zeng et al. (2024) analyzed the sources and health risks of fluoride in the Tarim, Xinjiang, China. Although these studies provide valuable insights into contamination status, they are limited in their ability to capture temporal variability in exposure risk, particularly in monitoring systems where fluoride is not routinely measured. In many river basins, including those in China, water quality monitoring programs prioritize a set of basic physicochemical parameters, while fluoride measurements are conducted only under extended or event-based monitoring schemes. As a result, substantial gaps remain in the continuous assessment of fluoride-related health risks, especially for vulnerable populations such as children.

Recent advances in machine learning offer new opportunities to address these limitations by leveraging nonlinear relationships among routinely monitored water quality parameters. Compared with traditional statistical methods, machine learning models are well suited for handling complex interactions and high-dimensional environmental datasets (Hossain et al., 2024; Li et al., 2025; Mohammadpour et al., 2024). Nevertheless, most existing applications of machine learning in fluoride research have focused on predicting fluoride concentrations rather than health-related risk metrics (Barzegar et al., 2017; Gawusu and Abu, 2025; Yang et al., 2024). Moreover, the lack of model interpretability has constrained the translation of data-driven predictions into decision-making processes informed by exposure and health considerations. Against this background, there is a critical need to shift from concentration-centered modeling to direct prediction of fluoride-related human health risks, particularly for children. Such an approach would align more closely with public health objectives and provide practical value under routine monitoring conditions. The integration of explainable machine learning techniques, such as SHAP (Shapley Additive Explanations), further enables the identification of key water quality indicators associated with elevated health risk, thereby enhancing the interpretability and credibility of predictive models.

In this study, we propose an explainable machine learning framework to predict fluoride-related human health risks in children using routinely monitored water quality parameters in the Huaihe River Basin, China. Specifically, multiple machine learning models were developed and compared to estimate children’s fluoride health risk indices, and the optimal model was interpreted using SHAP analysis to elucidate the contributions of individual predictors. The Huaihe River Basin was selected as the study area due to its dense population, intensive agricultural activities, and complex water quality conditions, which together create a representative scenario for fluoride health risks in large river basins. By directly linking routine water quality indicators to children’s fluoride health risk, this study aims to provide a practical tool for exposure screening and health-focused water quality management in regions with limited contaminant monitoring coverage.