Interpretable machine learning models to predict cadmium in wheat for safe production and soil management

Accurate prediction of cadmium (Cd) concentrations in wheat grain is essential for ensuring food safety and sustainable agriculture. Here, we developed a predictive model using nine machine learning (ML) algorithms based on a dataset of 1,339 soil-wheat grain pairs, with a focus on soil properties. The results showed that the eXtreme Gradient Boosting (XGBoost) model outperformed others, achieving superior predictive accuracy (R 2 = 0.90) compared to multiple linear regression (R 2 = 0.69). Through Shapley Additive Explanations (SHAP) analysis, soil total Cd (mean |SHAP| value, 0.20) and pH (0.08) were identified as key determinants, while soil Mn (0.06) and Zn (0.03) concentrations as minor determinants for wheat grain Cd. Soil Cd had a positive effect on grain Cd concentration, whereas soil pH, Mn and Zn showed negative effects. Extending the XGBoost model with 373 nation-scale paired data confirmed its robustness (R 2 = 0.86), and identified high-risk areas for Cd accumulation in southwest China and northwestern Henan province. An online application (https://wheat.cdpredict.cn) was developed for rapid Cd predictions in wheat. To ensure compliance with the wheat grain Cd limit of 0.1 mg/kg, soil Cd safety thresholds were established for different soil pH ranges. We further recommend that approximately 3.4% and 10.5% of cultivated soils should maintain Cd levels within 0.30 and 0.34 mg/kg, respectively. This interpretable ML model provides an actionable tool for managing soil contaminated with Cd to ensure the safe production of wheat.