NOVA Through Two Lenses: From Nutrients to Food Processing Classification in Finland's National Food Database

Ultra–processed food (UPF) intake has been linked to adverse health outcomes. However, NOVA processing groups are hard to assign consistently, and many national food composition databases do not include them. Nutrient–based classification models could fill this gap, but their performance across countries and databases is uncertain. FoodProX is a machine–learning classifier that predicts NOVA groups and a continuous food–processing score (FPro) from nutrient composition. To test whether such models transfer across databases, we retrained it on 182,344 mostly European Open Food Facts products. We evaluated it on 45,526 held-out Open Food Facts products and, without retraining, on 2,971 NOVA-labelled foods from a US database (FNDDS). We then applied it to the Finnish national food composition database (Fineli), which has no NOVA labels. Four–class accuracy was 0.855 on held–out products and 0.763 in FNDDS, with errors concentrated at the NOVA 3/4 boundary. UPFs were identified well, with precision–recall performance of 0.940 and 0.964, respectively (on a 0—1 scale), well above the chance levels of 0.463 and 0.711. On the same items, our FPro scores ranked foods similarly to those of the originally released US–trained FoodProX (Spearman ρ = 0.81 in FNDDS and 0.79 in Fineli), although our scores were lower on average. Sodium and sugars were the strongest predictors. Low fibre and iron also pushed predictions towards NOVA 4. In Fineli, the model assigned 55.9% of the 4,092 scored items to NOVA 4. We found no earlier NOVA or FPro annotations for this database. Nutrient–based models are most reliable for detecting ultra–processing and ranking processing intensity, and less reliable for assigning all four NOVA groups. Experts also agree poorly when assigning NOVA groups, so part of this difficulty may lie in the NOVA definitions themselves.