Efficient models for predicting durum wheat grain Cd conformity using soil variables and cultivars

[Display omitted] •Models were built to predict the compliance of durum wheat grain Cd content to regulatory thresholds.•Logistic regression and random forest were compared to multiple regression.•The models detected around 82% of the non-conformity cases with a reliability at around 75%.•Models were used to identify risky soil conditions for the phytoavailability of Cd.•Models and field trials showed a significant between-cultivar variability for the grain Cd content. Contamination of durum wheat grain by cadmium (Cd) threatens food safety and is of increasing concern because regulations concerning Cd are becoming stricter due to its toxicity. This work aimed at using soil variables and cultivar types to build models to predict whether durum wheat grain Cd will conform with current and possibly lower regulatory thresholds. We combined multiple Gaussian and logistic regressions and the random forest algorithm to take advantage of their strength. Models tested using cross-validation produced excellent performances including for the lowest regulatory threshold of 0.1 mg Cd/kg, half of the current one: 79–85% of the non-conformity cases were detected and the reliability of predictions was 69–82%. The models enabled identification of a x1.4 variability in grain Cd content between cultivars that do not have the low Cd accumulation allele of the Cdu1 gene. The models confirmed that for the grain Cd content, the between-cultivar variability had much less influence than the phytoavailability of Cd in soil, the critical contexts of which were characterized by the models. For farmers, these models are valuable tools to predict whether durum wheat production will conform with existing and future Cd regulation in foodstuffs.