We utilized GroceryDB, a dataset of 50,468 US products featuring FPro scores. Unlike the discrete NOVA classification, FPro uses a continuous 0–1 scale (0: unprocessed; 1: ultra-processed) to measure processing levels via the FoodProX model. After filtering for complete nutrition data and removing outliers, we trained a Random Forest on 12 log-transformed nutrients to predict FPro as both a continuous score and a four-class label.

This model was applied (without retraining) to CIQUAL 2025, the French national food-composition reference (3,484 foods). Predicted FPro classes were merged with Agribalyse 3.1 life-cycle assessment data to evaluate whether food processing levels predict environmental impact.