Application of machine learning to predict food processing level using Open Food Facts

Ultra-processed foods are increasingly linked to health issues like obesity, cardiovascular disease, type 2 diabetes, and mental health disorders due to poor nutritional quality. This first-of-its-kind study at such a scale uses machine learning to classify food processing levels (NOVA) based on the Open Food Facts dataset of over 900,000 products. Models including LightGBM, Random Forest, and CatBoost were trained on nutrient concentration data. LightGBM performed best, achieving 80-85% accuracy across different nutrient panels and effectively distinguishing minimally from ultra-processed foods. Exploratory analysis revealed strong associations between higher NOVA classes and lower Nutri-Scores, indicating poorer nutritional quality. Products in NOVA 3 and 4 also had higher carbon footprints and lower Eco-Scores, suggesting greater environmental impact. Allergen analysis identified gluten and milk as common in ultra-processed items, posing risks to sensitive individuals. Categories like Cakes and Snacks were dominant in higher NOVA classes, which also had more additives, highlighting the role of ingredient modification. This study, leveraging the largest dataset of NOVA-labeled products, emphasizes the health, environmental, and allergenic implications of food processing and showcases machine learning’s value in scalable classification. A user-friendly web tool is available for NOVA prediction using nutrient data: https://cosylab.iiitd.edu.in/foodlabel/.

💡 Research Summary

This study addresses the growing public‑health concern that ultra‑processed foods (UPFs) are linked to obesity, cardiovascular disease, type‑2 diabetes, mental‑health disorders, and environmental burdens. While the NOVA classification system is widely used to categorize foods by processing level, its manual labeling is labor‑intensive and suffers from low inter‑rater reliability. Leveraging the Open Food Facts (OFF) database—an open‑source repository containing detailed nutrient information for over 900,000 products—the authors develop an automated, nutrient‑based machine‑learning pipeline to predict NOVA classes (1–4).

Data preprocessing began by aligning OFF nutrients with the USDA Food and Nutrient Database for Dietary Studies (FNDDS) 2009‑10, identifying 44 common nutrients. For model training, the authors selected a subset of 7–8 nutrients with ≤15 % missing values. Missing data were handled via two strategies: simple mean imputation and a more sophisticated autoencoder (AEC) approach built in PyTorch. The autoencoder was tuned with Optuna, trained for 500 epochs, and used to generate plausible replacements for missing entries, thereby preserving latent nutritional patterns.

Three tree‑based algorithms—Random Forest (RF), LightGBM (LGBM), and CatBoost (CB)—were evaluated. Hyper‑parameter optimization employed RandomizedSearchCV, exploring extensive parameter spaces (e.g., number of leaves, learning rates, regularization terms). Because the dataset is heavily imbalanced (NOVA 1 and 4 dominate), the authors applied Synthetic Minority Over‑sampling Technique (SMOTE), undersampling, and class‑weight adjustments to mitigate bias. Performance metrics (accuracy, macro‑averaged F1, confusion matrices) show that LightGBM consistently outperformed the other models, achieving 80–85 % overall accuracy and macro‑F1 scores between 0.78 and 0.81. The model excelled at distinguishing minimally processed (NOVA 1) from ultra‑processed (NOVA 4) items, while still maintaining reasonable discrimination among intermediate classes.

Beyond predictive modeling, the paper conducts an extensive exploratory data analysis (EDA). Categorical relationships—between NOVA class and product category, allergen presence, Nutri‑Score grade, Eco‑Score grade—were examined using chi‑square tests and Cramér’s V to assess association strength. Continuous variables (carbon footprint, Nutri‑Score numeric values, Eco‑Score) were compared across NOVA levels with Kruskal‑Wallis tests. Findings reveal a clear trend: higher NOVA classes correlate with poorer Nutri‑Score (more “E” grades), lower Eco‑Score (more “C” grades), and higher carbon footprints, indicating that ultra‑processed items are nutritionally inferior and environmentally more demanding. Allergen analysis shows gluten and milk as the most frequent allergens in NOVA 4 products, underscoring heightened risk for sensitive consumers.

Network analysis of product categories, using Louvain community detection on a co‑occurrence graph of the top 50 categories per NOVA class, visualizes clusters such as cakes, snacks, and sugary beverages that concentrate in the ultra‑processed segment. This structural insight can guide policymakers and industry stakeholders in targeting reformulation or labeling interventions.

The authors also compare their work to prior efforts that relied on small datasets (≈3,000 items) and primarily natural‑language processing of ingredient lists. By scaling to ~800,000 items and focusing solely on nutrient profiles, the study demonstrates that macro‑nutrient patterns—shaped by industrial processing—are sufficient for robust NOVA classification. The inclusion of autoencoder‑based imputation further improves data quality without manual curation.

To translate research into practice, the team released a user‑friendly web application (https://cosylab.iiitd.edu.in/foodlabel/) where users can input nutrient values and receive instant NOVA predictions. This tool offers immediate utility for food manufacturers, regulators, and consumers seeking rapid, data‑driven assessments of processing level.

In conclusion, the paper validates that machine‑learning models trained on a limited set of standardized nutrient features can accurately predict food processing categories at a massive scale. It also provides compelling evidence that ultra‑processed foods are associated with lower nutritional quality, higher environmental impact, and greater allergen prevalence. Future directions include integrating additional data modalities (e.g., ingredient text, packaging images), refining regional models to account for dietary patterns, and deploying the system in real‑time labeling pipelines to support healthier food environments worldwide.