NOVA Food Processing System Insights

• NOVA Food Processing System is a classification framework that categorizes foods into four groups based on processing intensity and ingredient composition.
• It leverages large-scale databases and machine learning to improve assignment reproducibility and scalability in nutritional epidemiology.
• Advanced NLP and ensemble models integrate nutrient and ingredient data, offering interpretable, continuous scoring of food processing levels.

The NOVA Food Processing System is an ontological and computational framework for the systematic classification of foods according to the extent and purpose of their processing. Initially advanced by Monteiro and colleagues, NOVA has become the dominant paradigm in nutritional epidemiology for associating food processing with a spectrum of health, environmental, and regulatory outcomes. Recent informatics advances have further operationalized NOVA using machine learning and natural language processing, providing scalable and reproducible alternatives to manual assignment.

1. NOVA Classification System: Definitions, Criteria, and Controversies

NOVA categorizes foods into four mutually exclusive classes based on industrial processing practices:

Group	Definition	Representative Examples
1	Unprocessed or minimally processed foods: physical or simple processes such as cleaning, drying, roasting, pasteurization; no addition of exogenous substances.	Fresh fruits/vegetables, plain milk, eggs, raw nuts, legumes
2	Processed culinary ingredients: substances extracted from group 1 via pressing, milling, refining, or other fractionation; used exclusively as food ingredients.	Salt, sugar, oils, butter, vinegar, wheat flour
3	Processed foods: items produced by adding group 2 ingredients (e.g., salt, sugars, oils) to group 1, typically with 2–3 components to enhance shelf-life or palatability.	Canned vegetables, bread, cheese, fruit in syrup, smoked meats
4	Ultra-processed foods (UPF): industrial formulations with five or more ingredients, often with additives (emulsifiers, colorants, flavor enhancers, fortificants).	Soft drinks, packaged snacks, instant noodles, confectioneries, breakfast cereals

Assignment is primarily based on the extent and purpose of processing, using ingredient "deformulation" (reverse-engineering from labels) when direct manufacturing details are unavailable (Ispirova et al., 20 May 2025). Key criteria for UPF include the presence of industrial additives, fractionated or reconstituted ingredients, and fortification agents.

Several limitations are recognized:

• Subjectivity in interpreting ingredient lists and industrial processes
• Poor inter-rater reliability and frequent revision of definitions (eight discrete updates between 2009–2017)
• Manual assignment’s low scalability for epidemiological datasets and public databases
• Coarse granularity, often collapsing nutritionally distinct items within a single group (e.g., whole-grain vs. sugary cereals)
• Absence of quantitative nutrient thresholds for group demarcation (Ispirova et al., 20 May 2025).

2. Data Resources and Preprocessing Protocols

Large-scale application of the NOVA system has leveraged structured food composition databases:

• Open Food Facts (OFF): ~900,000 products worldwide, annotated with nutrient panels, category labels, allergens, Nutri-Score, and Eco-Score. Nutrient mapping to USDA FNDDS 2009–2010 enables harmonized feature sets, varying from "44-nutrient" down to "7-nutrient" panels (Arora et al., 19 Dec 2025).
• FNDDS (USDA): 2,970 items with up to 102 nutrient measurements per food, manually labeled for NOVA classes and used in model training (Arora et al., 2024).
• Open Food Facts (OFF) English subset: 149,960 products with complete NOVA label, ingredient list, and 11-nutrient panel used for multimodal AI and model benchmarking (Ispirova et al., 20 May 2025).

Preprocessing includes alignment of nutrient codes, z-score normalization

$z_i = \frac{x_i - \mu_i}{\sigma_i}$

for all features, and imputation strategies (mean imputation and autoencoder-based reconstruction) to address missing data, especially outside sparse core panels (Arora et al., 19 Dec 2025).

3. Computational Modeling and Algorithms

Emerging informatics solutions augment manual NOVA assignment with machine learning and feature engineering:

• Ensemble Trees: LightGBM (histogram gradient boosting), Random Forest, CatBoost, and Gradient Boost are consistently evaluated, with hyperparameters tuned via RandomizedSearchCV over tree count, depth, learning rate, and regularization terms (Arora et al., 19 Dec 2025, Arora et al., 2024). Class imbalance (notably NOVA 2, which is underrepresented) is mitigated via SMOTE oversampling combined with stratified k-fold cross-validation.
• Multimodal and NLP Models: Transformer-based text encoders (BERT, BioBERT, DistilBERT, XLM-RoBERTa, GPT-2) extract dense representations from ingredient lists and food descriptions, concatenated with nutrient features for downstream classification. The [CLS] token’s final hidden state encodes the sample for input to ensemble classifiers (Ispirova et al., 20 May 2025, Arora et al., 2024).
• Continuous Scoring (FPro): The FoodProX model computes a probability vector over NOVA classes and projects it onto a continuous [0,1] axis:

$FPro(x) = \frac{1 - p_{1}(x) + p_{4}(x)}{2}$

to capture gradations from minimally to ultra-processed beyond categorical labels (Ispirova et al., 20 May 2025).

4. Model Evaluation, Feature Attribution, and Interpretability

Performance across datasets and panels is consistently quantified using macro-averaged accuracy, F1-score, Matthews Correlation Coefficient (MCC), area under the ROC curve (AUC), and area under the precision-recall curve (AUP). Typical results on the Open Food Facts dataset and FNDDS panels:

Nutrient Panel	Best Model	Macro F1	MCC	AUC (NOVA 1/4)
7–8 OFF core	LightGBM	0.84	0.69	--
13 FDA/FNDDS	Gradient Boost	0.9284	0.8425	--
102 FNDDS	LGBM	0.9411	0.8691	--
12 basic (FoodProX)	RF (FNDDS)	--	--	0.98/0.98
13 + GPT-2 NLP	LGBM	0.9583	0.9091	--
11 nutrients + BERT	XGBoost (OFF)	--	--	0.995/0.992

SHAP (SHapley Additive exPlanations) values and feature attribution highlight sodium, total fat, sugars, energy, and presence of folic acid fortification as principal determinants, with systematic shifts in sodium and sugar concentration distinguishing higher NOVA classes (Arora et al., 19 Dec 2025, Arora et al., 2024). Inclusion of NLP embeddings improves macro F1 to ~0.95–0.96 in recent architectures (Arora et al., 2024, Ispirova et al., 20 May 2025).

5. Exploratory Analyses: Nutrition Quality, Environmental, and Allergen Correlations

Data-driven classification reveals significant associations between NOVA classes and nutritional quality (Nutri-Score), environmental scores (Eco-Score), and allergenic burden.

• Nutri-Score: NOVA 1 products cluster in A/B, while 33.11% of NOVA 4 foods are D. Kruskal–Wallis H=127,986, p<10^–5; Pearson r(NOVA, Nutri-Score)=0.40 (Arora et al., 19 Dec 2025).
• Eco-Score and Carbon Footprint: NOVA 1 is skewed toward favorable Eco-Score (A/B), while NOVA 4 is overrepresented in C/D. Median carbon footprints increase modestly with processing (r=0.12) (Arora et al., 19 Dec 2025).
• Allergens: Ultra-processed (NOVA 4) foods contain higher proportions of milk (34%), gluten (25%), nuts (14%), and peanuts (~9%), and average 1.3 allergenic ingredients per product, compared to 0.4 in minimally processed foods (Arora et al., 19 Dec 2025).
• Additives: Strong positive correlation between additive count and NOVA class (r=0.42), particularly in categories like cakes, sweets, and prepared meals, using community detection on food-category networks (Arora et al., 19 Dec 2025).

6. Web-Based and Scalable Applications

Several open-access web tools operationalize NOVA classification for end-users and researchers:

• FoodLabel NOVA predictor (OFF): https://cosylab.iiitd.edu.in/foodlabel/ Users input per-100g values for a 7-nutrient panel; model returns NOVA class with class-probabilities and SHAP explanations.
• Food Processing Level Predictor (FNDDS): https://cosylab.iiitd.edu.in/food-processing/ Accepts 13/65/102 nutrient panels and optional text fields, running ensemble and NLP-augmented models for classification, probability, and interpretability output (Arora et al., 19 Dec 2025, Arora et al., 2024).

Backend APIs perform feature scaling, generate embeddings if needed, and serve predictions using pickled models. These platforms enable rapid, large-scale screening of food products, supporting research on inventories, dietary surveys, policy analytics, and consumer guidance.

7. Implications and Future Directions

Automated NOVA classification minimizes subjectivity and improves reproducibility, facilitating epidemiological research and regulatory efforts at scale (Ispirova et al., 20 May 2025). Continuous scoring (FPro), flexible use of missing data via transformer-based models, and integration of structured (nutrient) and unstructured (textual) data expand the applicability of NOVA to heterogeneous, global food supply chains. Use-cases include epidemiology, public-health nutrition, environmental assessment, regulatory monitoring, and digital health applications. Ongoing development addresses limitations related to granularity, ambiguous ingredient labeling, and shifting patterns of food processing and fortification.

Recent advances demonstrate that a small number of nutrient features, used as input to robust tree ensembles and multimodal AI, recover NOVA designations with >90% macro F1 and MCC on large real-world datasets. SHAP analysis supports mechanistic understanding, and continuous scoring refines dietary guidance beyond categorical bins (Arora et al., 19 Dec 2025, Arora et al., 2024, Ispirova et al., 20 May 2025). A plausible implication is the convergence of informatics and public health nutrition in the development of evidence-based, reproducible food classification pipelines supporting population-scale dietary research and consumer-facing digital tools.