Predictive Modeling of Food Processing Levels

• Predictive modeling of food processing levels is a computational method that applies machine learning and NLP to classify foods by the degree of processing using the NOVA framework.
• It employs standardized datasets like Open Food Facts and FNDDS with rigorous preprocessing and ensemble models to achieve high prediction accuracy.
• The approach leverages multimodal data fusion and interpretability tools such as SHAP to enhance transparency and guide actionable dietary recommendations.

Predictive modeling of food processing levels refers to the computational task of inferring the degree of industrial processing applied to food products—most commonly using the NOVA classification system—through structured and unstructured data such as nutrient profiles, ingredient lists, and textual product metadata. Advances in machine learning, natural language processing, and multimodal data fusion have enabled scalable, reproducible, and interpretable models for this purpose, with direct applications in public health, food science, epidemiology, and consumer informatics.

1. Foundations of Food Processing Level Classification

The primary framework for categorizing food processing in computational models is the NOVA classification, which stratifies foods into four ordinal categories:

• NOVA 1: Unprocessed or minimally processed foods
• NOVA 2: Processed culinary ingredients
• NOVA 3: Processed foods
• NOVA 4: Ultra-processed foods

Traditional assignments relied on manual inspection of ingredient lists and processing technologies, introducing subjectivity and limited scalability. Data-driven approaches seek to overcome these limitations by formalizing the mapping from objectively measured features (nutrient concentrations, ingredient counts, or text) to NOVA classes using supervised learning (Arora et al., 19 Dec 2025, Arora et al., 2024, Ispirova et al., 20 May 2025).

2. Datasets, Features, and Preprocessing

Predictive modeling studies leverage large, curated databases such as Open Food Facts (OFF) and the USDA Food and Nutrient Database for Dietary Studies (FNDDS) (Arora et al., 19 Dec 2025, Arora et al., 2024, Ispirova et al., 20 May 2025). These repositories provide:

• Per-product nutrient composition vectors (varying from 7 to 102 nutrients).
• Ingredient lists, food category labels, and free-text fields.

Key preprocessing steps include:

• Alignment of nutrient features across datasets (e.g., mapping OFF nutrients to FNDDS standards, identifying common panels such as the 13-feature FDA Nutrition Facts).
• Treatment of missing values: mean imputation and autoencoder-based imputation for large panels; dropping products/columns for reduced panels (Arora et al., 19 Dec 2025).
• Z-score normalization or standard scaling, particularly relevant for neural models though less critical for tree ensembles.

Class imbalance, typically due to over-representation of ultra-processed foods, is addressed via SMOTE or stratified sampling during cross-validation (Arora et al., 2024).

3. Machine Learning and NLP Approaches

The principal supervised models are ensemble-based classifiers: Random Forest, LightGBM (gradient boosting decision trees), CatBoost, and variants of GradientBoost and XGBoost (Arora et al., 19 Dec 2025, Arora et al., 2024, Ispirova et al., 20 May 2025). Hyperparameters are tuned via randomized or grid search under k-fold cross-validation regimes.

Nutrient composition models operate on input vectors of varying length:

Panel	# Features	Best Model	F1-score	MCC
Large	44–102	LightGBM	0.94	0.87
Medium	8–65	Random Forest/LGBM	0.93–0.94	0.85
Reduced	7–13	GradientBoost/GBDT	0.84–0.93	0.69–0.84

For food descriptions and ingredient lists, transformer-based LLMs (BERT, DistilBERT, XLM-RoBERTa, BioBERT, GPT-2) extract dense embeddings (typically the 768-dimensional [CLS] token) that encode semantic properties of foods (Arora et al., 2024, Ispirova et al., 20 May 2025). These embeddings, optionally concatenated with numeric nutrient features, are further classified using tree ensembles or shallow neural networks.

Multimodal models combine predictions from nutrient-only, text-embedding, and additive-count pipelines, often averaging probabilities to achieve increased robustness (Ispirova et al., 20 May 2025).

4. Model Training, Evaluation, and Interpretability

Training paradigms emphasize reproducibility and methodological rigor:

• Stratified k-fold cross-validation (typically k=5 or 10).
• Hyperparameter grids tailored per model and feature panel.
• Performance metrics: accuracy, precision, recall, F1-score, Matthews Correlation Coefficient (MCC); ROC AUC and area under the precision-recall curve (AUP) are reported for NOVA class discrimination (Arora et al., 19 Dec 2025, Arora et al., 2024, Ispirova et al., 20 May 2025).

Explanatory analyses employ SHAP (SHapley Additive exPlanations) to quantify nutrient or feature contributions to model predictions. Across datasets, sugars, sodium, total fat, and carbohydrate are primary discriminants of ultra-processing, with lower sodium and fat indicative of minimal processing.

Class-wise performance trends reveal highest precision and recall for ultra-processed (NOVA 4) and processed foods (NOVA 3), and reduced but still acceptable discrimination for minimally processed (NOVA 1) and culinary ingredient (NOVA 2) classes (Arora et al., 19 Dec 2025).

5. Continuous Scoring and Model Output Enhancement

To capture within-class heterogeneity, the FoodProX system (Random Forest-based) introduces the FPro continuous processing score:

$FPro_k = \frac{1-p_1^{(k)}+p_4^{(k)}}{2}$

where $p_i^{(k)}$ is the probability assigned by the classifier to NOVA class $i$ for sample $k$. $FPro_k$ provides a scalar measure on $[0,1]$ interpolating between minimally and ultra-processed (Ispirova et al., 20 May 2025). This continuous output supports personalized dietary recommendations, food environment assessment, and epidemiological research.

6. Integrative Correlation and Exploratory Analyses

Beyond strict classification, studies systematically investigate the phenotypic correlates of processing level:

• Strong association of higher NOVA class with higher number of additives and lower nutritional quality (as measured by Nutri-Score).
• Moderate to weak association with environmental impact metrics such as carbon footprint and Eco-Score.
• Allergen risk: milk and gluten are prevalent especially in ultra-processed categories (Arora et al., 19 Dec 2025).
• Unstructured text features (food category, description) provide orthogonal signals for classifying items where nutrient or ingredient data are sparse.

Chi-square, Cramér’s V, and Kruskal–Wallis tests corroborate and quantify these associations, facilitating multidimensional public health risk assessments based on predicted processing levels (Arora et al., 19 Dec 2025).

7. Deployment, Reproducibility, and Best Practices

Operationalization of predictive models is achieved via open-access web servers (e.g., https://cosylab.iiitd.edu.in/foodlabel/ and https://cosylab.iiitd.edu.in/food-processing/), enabling real-time classification from user-entered nutrient panels. Backend implementations rely on pre-trained LightGBM, Random Forest, and GradientBoost pipelines; Flask-based web applications handle inference and display.

Best practices include:

• Use of large, standardized food composition datasets (FNDDS, OFF).
• Harmonization of ingredient ontologies.
• Rigorous documentation of preprocessing and hyperparameter sets.
• Sharing of code, splits, and environments to ensure inter-study reproducibility.
• Reporting of both discrete and continuous processing predictions to maximize downstream scientific utility (Ispirova et al., 20 May 2025, Arora et al., 2024, Arora et al., 19 Dec 2025).

Ensemble and multimodal strategies are recommended where data extensiveness or structural heterogeneity justify additional modeling complexity.

---

Predictive modeling of food processing levels—anchored in the NOVA framework and advanced by machine learning, NLP, and data fusion—has demonstrated robustness and scalability across national and global food supply databases. Key nutrients (sugars, sodium, fat) and ingredient/additive signals consistently emerge as primary discriminators, and state-of-the-art models now achieve F1-scores of 0.84–0.96 depending on feature richness. As the field evolves, multimodal semantic models and continuous scoring systems such as FPro are poised to further enhance the resolution and actionable value of food processing classification (Arora et al., 19 Dec 2025, Arora et al., 2024, Ispirova et al., 20 May 2025).