Robustness of nutrient-based NOVA food classification under profile-separated nested validation

Introduction Nutrient profiles contain information associated with food processing, but estimates of automated NOVA classification can be distorted by ambiguous record linkage, severe class imbalance, duplicate profiles, and non-nested model selection. This study evaluates how well four-class NOVA labels can be recovered from nutrient composition after addressing these sources of optimism. Methods The analysis used 2,484 manually labeled foods from the USDA Food and Nutrient Database for Dietary Studies, represented by 99 nutrient variables. Identical nutrient profiles were assigned to the same fold, and logistic regression, a radial-basis support vector machine, random forest, and histogram gradient boosting were compared with a prior-probability baseline using three repeats of five-fold nested stratified group cross-validation. Hyperparameters and all transformations were selected within each outer training partition. Performance was summarized using class-balanced and chance-corrected metrics, corrected repeated-cross-validation tests, and fold-stable multiclass Shapley attribution. Results Histogram gradient boosting achieved accuracy 0.921 and macro-F1 0.814 (SD 0.029; 95% fold interval 0.798–0.831). Random forest reached macro-F1 0.788; the mean difference of 0.026 was not statistically reliable after repeated-fold correction (Holm-adjusted p = 0.203). Class-specific F1 was 0.639 for NOVA 2 and 0.958 for NOVA 4. Discussion Nutrient measurements therefore provide a strong but incomplete correlate of NOVA class. The models are suitable for methodological screening within this data source, but they do not replace ingredient- and process-based expert classification; external validation on contemporary branded foods and independently adjudicated labels remains necessary.