FoodProX: Nutrient-Based Processing Inference

• FoodProX is a machine-learning approach that uses nutrient quantities to predict NOVA classes and derive a continuous FPro score.
• It addresses limitations of expert-driven frameworks by capturing multivariate nutrient changes through a probabilistic, reproducible model.
• Validation on databases like FNDDS and Open Food Facts shows high performance, supporting its use in nutritional epidemiology and public health surveillance.

Searching arXiv for FoodProX and closely related food-processing prediction work. FoodProX is a machine-learning approach for inferring food processing level from nutrient composition, introduced as a multi-class random forest classifier that predicts a probability distribution over the four NOVA classes and, from that distribution, a continuous processing score called FPro (Ispirova et al., 20 May 2025). It was formulated to address recurrent limitations of expert-driven processing frameworks—especially the subjectivity, evolving definitions, and reproducibility problems associated with NOVA—by using nutrient quantities in grams per 100 g of food, a data modality described as consistently regulated and widely available worldwide. In the literature summarized here, FoodProX occupies a specific niche within food informatics: it models processing as a multivariate compositional signal rather than as a purely ingredient-list heuristic, an image-recognition task, or a direct laboratory assay (Ispirova et al., 20 May 2025).

1. Conceptual origin and scientific rationale

FoodProX emerged from a critique of traditional food-processing frameworks. NOVA is treated as influential in epidemiology but limited by expert dependence, incomplete information, and low reproducibility; Nutri-Score is described as easier to apply but not a processing measure; SIGA is more ingredient-aware but less portable to datasets that contain nutrient panels without detailed branded-product ingredient data (Ispirova et al., 20 May 2025). FoodProX was proposed to operate precisely in that gap.

Its central premise is that food processing perturbs coordinated nutrient composition in systematic ways. The chapter situates this in a broader “foodome” perspective: nutrient concentrations follow broad physiological and physicochemical regularities, and industrial processing shifts those regularities through removal, enrichment, fortification, dilution, and the addition of sodium, sugars, and fats (Ispirova et al., 20 May 2025). FoodProX is therefore designed not to detect a single nutrient threshold, but to recognize multivariate deviations across the nutrient profile.

This design makes FoodProX compatible with NOVA while not being reducible to manual NOVA rules. The model is supervised on NOVA labels, so it inherits NOVA’s ontology, but its inference mechanism is statistical rather than expert-coded. A plausible implication is that FoodProX should be understood less as a replacement for all processing taxonomies than as a reproducible inferential layer over nutrient databases.

2. Formal definition of FoodProX and FPro

In the original formulation reported in the chapter, FoodProX is a multi-class random forest classifier trained on nutrient quantities in grams per 100 g of food to predict membership in the four NOVA classes: NOVA 1, NOVA 2, NOVA 3, and NOVA 4 (Ispirova et al., 20 May 2025). For each food $k$, the model outputs a class-probability vector

$$\mathbf{p}^{(k)} = \left(p_1^{(k)}, p_2^{(k)}, p_3^{(k)}, p_4^{(k)}\right), \qquad \sum_{i=1}^4 p_i^{(k)} = 1, \qquad p_i^{(k)} \ge 0.$$

The discrete class prediction is the argmax over that vector:

$$\hat{y}^{(k)} = \arg\max_{i \in \{1,2,3,4\}} p_i^{(k)}.$$

The distinctive contribution is FPro, a continuous processing score derived from the probability vector rather than learned as a separate target. The chapter defines FPro as the orthogonal projection of the probability vector onto the line connecting the NOVA 1 vertex $(1,0,0,0)$ and the NOVA 4 vertex $(0,0,0,1)$, yielding

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

This construction makes $FPro \approx 0$ correspond to minimally processed or NOVA 1-like foods, and $FPro \approx 1$ correspond to ultra-processed or NOVA 4-like foods (Ispirova et al., 20 May 2025). The chapter describes the probability vector geometrically as a point in the simplex, visualized for four classes as a tetrahedron, and FPro as a projection that preserves ambiguity rather than forcing a hard categorical decision.

That distinction between hard class and continuous score is foundational. FoodProX is the classifier; FPro is the continuous processing gradient computed from its probabilistic output. The chapter further states that FPro is architecture-agnostic: any model that outputs probabilities over the four NOVA classes can, in principle, be used to compute it (Ispirova et al., 20 May 2025).

3. Training data, feature regimes, and implementation settings

The original FoodProX training resource was USDA FNDDS 2009–2010. The chapter reports that FNDDS contains thousands of model foods, has no missing values for the nutrient profiles used there, and served as the basis for five-fold stratified cross-validation (Ispirova et al., 20 May 2025). Figure 1 in the chapter mentions 4,889 foods reported in NHANES 2009–2010 data, and 2,484 items, or 34.25%, had a unique NOVA class assignment; the remainder were left unclassified or required decomposition into constituent ingredients (Ispirova et al., 20 May 2025). This is important because FoodProX learns from manual NOVA labels, but the chapter is explicit that those labels were available only for a subset with sufficiently unambiguous assignment.

Three nutrient panels were used in the original implementation: 99 nutrients, 62 nutrients, and 12 nutrients (Ispirova et al., 20 May 2025). The text frames these as progressively reduced data-availability regimes, ranging from rich epidemiological databases to more limited labeling scenarios. The chapter does not report random forest hyperparameters, explicit class reweighting, or any probability calibration method, and it does not state that nutrient features were log-transformed before model fitting.

The same chapter later reports FoodProX-style random-forest models on Open Food Facts. In that case study, the dataset was filtered to 149,960 products with English names and complete information for product name, ingredient list, NOVA classification, and an 11-nutrient panel consisting of proteins, fat, carbohydrates, sugars, fiber, calcium, iron, sodium, cholesterol, saturated fat, and trans-fat (Ispirova et al., 20 May 2025). Two tabular variants were used: one with those 11 nutrients and one with the same panel plus total number of additives. The validation protocol there comprised a 20% stratified holdout set for hyperparameter tuning and the remaining 80% split into five fixed train/test stratified folds.

A separate large-scale Open Food Facts benchmark later extended the nutrient-only NOVA prediction paradigm to substantially larger datasets, using LightGBM, Random Forest, and CatBoost rather than the original FoodProX random forest (Arora et al., 19 Dec 2025). That study used 875,075 products in one experiment, 681,950 in a 7-nutrient experiment, and 479,090 in an 8-nutrient experiment. For the 44-nutrient setting it compared mean imputation with autoencoder-based imputation, and reported that autoencoder imputation performed notably worse than mean imputation (Arora et al., 19 Dec 2025). Although that paper does not itself define FPro, it is directly adjacent to the FoodProX program because it tests the same core hypothesis at much larger scale: nutrient profiles contain sufficient signal to recover NOVA processing level.

4. Validation evidence and quantitative performance

The validation evidence for FoodProX is strongest in two domains: internal cross-validation on FNDDS and transfer to large branded-product databases. The chapter reports that the original 12-nutrient FoodProX model achieved the following AUC values under five-fold stratified cross-validation on FNDDS: $0.9804 \pm 0.0012$ for NOVA 1, $0.9632 \pm 0.0024$ for NOVA 2, $$\mathbf{p}^{(k)} = \left(p_1^{(k)}, p_2^{(k)}, p_3^{(k)}, p_4^{(k)}\right), \qquad \sum_{i=1}^4 p_i^{(k)} = 1, \qquad p_i^{(k)} \ge 0.$$0 for NOVA 3, and $$\mathbf{p}^{(k)} = \left(p_1^{(k)}, p_2^{(k)}, p_3^{(k)}, p_4^{(k)}\right), \qquad \sum_{i=1}^4 p_i^{(k)} = 1, \qquad p_i^{(k)} \ge 0.$$1 for NOVA 4 (Ispirova et al., 20 May 2025). The corresponding AUP values were $$\mathbf{p}^{(k)} = \left(p_1^{(k)}, p_2^{(k)}, p_3^{(k)}, p_4^{(k)}\right), \qquad \sum_{i=1}^4 p_i^{(k)} = 1, \qquad p_i^{(k)} \ge 0.$$2, $$\mathbf{p}^{(k)} = \left(p_1^{(k)}, p_2^{(k)}, p_3^{(k)}, p_4^{(k)}\right), \qquad \sum_{i=1}^4 p_i^{(k)} = 1, \qquad p_i^{(k)} \ge 0.$$3, $$\mathbf{p}^{(k)} = \left(p_1^{(k)}, p_2^{(k)}, p_3^{(k)}, p_4^{(k)}\right), \qquad \sum_{i=1}^4 p_i^{(k)} = 1, \qquad p_i^{(k)} \ge 0.$$4, and $$\mathbf{p}^{(k)} = \left(p_1^{(k)}, p_2^{(k)}, p_3^{(k)}, p_4^{(k)}\right), \qquad \sum_{i=1}^4 p_i^{(k)} = 1, \qquad p_i^{(k)} \ge 0.$$5. The chapter emphasizes that these values are notable because NOVA is not explicitly defined in nutrient terms.

The Open Food Facts case study in the same chapter showed that nutrient-only FoodProX-style models remained strong on real-world branded products. With 11 nutrients alone, the reported AUC values were 0.988, 0.983, 0.926, and 0.948 for NOVA 1–4, respectively; with 11 nutrients plus additive count, they increased to 0.993, 0.988, 0.966, and 0.980 (Ispirova et al., 20 May 2025). The AUP values also improved, especially for NOVA 3 and NOVA 4, indicating that additive count acts as a low-cost augmentation for nutrient-based prediction when such metadata are available.

The larger Open Food Facts study provides a complementary benchmark rather than a direct reimplementation of FoodProX. Its best overall model was LightGBM on an 8-nutrient panel, with accuracy $$\mathbf{p}^{(k)} = \left(p_1^{(k)}, p_2^{(k)}, p_3^{(k)}, p_4^{(k)}\right), \qquad \sum_{i=1}^4 p_i^{(k)} = 1, \qquad p_i^{(k)} \ge 0.$$6, F1 $$\mathbf{p}^{(k)} = \left(p_1^{(k)}, p_2^{(k)}, p_3^{(k)}, p_4^{(k)}\right), \qquad \sum_{i=1}^4 p_i^{(k)} = 1, \qquad p_i^{(k)} \ge 0.$$7, and MCC $$\mathbf{p}^{(k)} = \left(p_1^{(k)}, p_2^{(k)}, p_3^{(k)}, p_4^{(k)}\right), \qquad \sum_{i=1}^4 p_i^{(k)} = 1, \qquad p_i^{(k)} \ge 0.$$8 (Arora et al., 19 Dec 2025). External validation on 2,970 FNDDS 2009–10 foods yielded accuracies of 81.2% for the best 7-nutrient model, 80.7% for the best 8-nutrient model, and 77.8% for the best 44-nutrient model. This result is methodologically significant because it suggests that compact, relatively complete nutrient panels can outperform broader but sparse nutrient spaces after imputation.

The following table summarizes the main quantitative regimes reported across the literature.

Setting	Inputs	Representative results
Original FoodProX on FNDDS	Random forest; 99/62/12 nutrient panels	12-nutrient AUC: 0.9804, 0.9632, 0.9696, 0.9789 for NOVA 1–4
FoodProX-style OFF case study	Random forest; 11 nutrients or 11 nutrients + additive count	AUC with 11 nutrients + additives: 0.993, 0.988, 0.966, 0.980
Large-scale OFF benchmark	LightGBM/RF/CatBoost; 7, 8, or 44 nutrients	Best: LightGBM with 8 nutrients, accuracy 0.85, F1 0.84, MCC 0.69

Taken together, these studies support a consistent empirical claim: nutrient composition alone carries substantial information about processing class, particularly at the extremes of the NOVA continuum. They also show that intermediate classes, especially NOVA 2 and NOVA 3, remain harder than NOVA 1 and NOVA 4 (Ispirova et al., 20 May 2025).

5. Interpretation, examples, and within-class heterogeneity

FoodProX’s interpretive value lies less in formal feature-attribution analyses than in the behavior of its probability vector and FPro ranking. The chapter’s canonical example is the onion trajectory: raw onion has $$\mathbf{p}^{(k)} = \left(p_1^{(k)}, p_2^{(k)}, p_3^{(k)}, p_4^{(k)}\right), \qquad \sum_{i=1}^4 p_i^{(k)} = 1, \qquad p_i^{(k)} \ge 0.$$9, whereas onion rings have $$\hat{y}^{(k)} = \arg\max_{i \in \{1,2,3,4\}} p_i^{(k)}.$$0 (Ispirova et al., 20 May 2025). The argument is that this shift reflects coordinated nutrient change rather than a single ingredient flag. The text notes that about 75% of nutrients change by more than 10% between raw onion and onion rings, and more than half change tenfold.

A second illustration involves breakfast cereals, all manually assigned to NOVA 4 but separated by FPro. The reported values are 0.5658 for Post Shredded Wheat’n Bran, 0.5685 for Post Shredded Wheat, 0.9603 for Post Grape-Nuts, and 0.9999 for Post Honey Bunches of Oats with Almonds (Ispirova et al., 20 May 2025). This is a decisive example of the difference between discrete NOVA labeling and continuous FoodProX scoring: foods that are all “ultra-processed” in categorical terms can nevertheless occupy very different locations on a nutrient-derived processing continuum.

The later large-scale Open Food Facts benchmark provides more explicit feature interpretation through SHAP. There, sodium and total fat were most influential for NOVA 1, total fat and energy for NOVA 2, sugars total and carbohydrate for NOVA 3, and sugars total and sodium for NOVA 4 (Arora et al., 19 Dec 2025). The paper interprets these as stable nutrient signatures of increasing industrial formulation. In combination with the earlier FoodProX framing, this suggests that FPro captures a multivariate pattern of sodium enrichment, sugar enrichment, fat enrichment, fortification, and depletion or restructuring of other nutrients rather than a single mechanistic definition of processing.

6. Extensions, uses, and limitations

FoodProX has been proposed for several applications. In nutritional epidemiology, it can standardize food-processing measurement in large nutrient databases and reduce inter-rater variability relative to manual NOVA coding (Ispirova et al., 20 May 2025). In public-health surveillance, it can classify foods at scale across national food databases and retail inventories. The chapter also frames FPro as useful for product reformulation analysis, category-level ranking of less processed alternatives, and large-scale grocery database studies.

The same chapter extends the FoodProX paradigm beyond tabular nutrient inputs. In its Open Food Facts case study, BERT and BioBERT were used to embed textual descriptions built from food name, ingredient list, and nutrients serialized into sentences of the form “$$\hat{y}^{(k)} = \arg\max_{i \in \{1,2,3,4\}} p_i^{(k)}.$$1 has the ingredients: $$\hat{y}^{(k)} = \arg\max_{i \in \{1,2,3,4\}} p_i^{(k)}.$$2, and the nutrients: $$\hat{y}^{(k)} = \arg\max_{i \in \{1,2,3,4\}} p_i^{(k)}.$$3, $$\hat{y}^{(k)} = \arg\max_{i \in \{1,2,3,4\}} p_i^{(k)}.$$4, …” (Ispirova et al., 20 May 2025). The embedding extracted was the 768-dimensional $$\hat{y}^{(k)} = \arg\max_{i \in \{1,2,3,4\}} p_i^{(k)}.$$5 token, and the resulting classifiers could also be used to compute FPro because the score is architecture-agnostic. This places FoodProX within a broader multimodal shift in food informatics: nutrient-only random forests remain efficient and relatively interpretable, while language-model pipelines can exploit noisy, heterogeneous branded-product metadata.

Its limitations are equally central. FoodProX is supervised on NOVA labels and therefore inherits NOVA’s ambiguities, inconsistencies, and evolving definitions (Ispirova et al., 20 May 2025). Nutrient profiles do not directly encode additives not reflected on the panel, food-matrix disruption, extrusion, emulsification, packaging contaminants, or specific industrial techniques. Class imbalance, especially the underrepresentation of NOVA 2, remains a recurrent issue in both the chapter and the later large-scale Open Food Facts benchmark (Ispirova et al., 20 May 2025). Database quality also matters: Open Food Facts is described as containing formatting errors, implausible nutrient values, possible NOVA mislabels, and inconsistent ingredient data.

FoodProX is therefore distinct from other branches of food AI. It does not perform prepared-dish recognition from images, 3D portion estimation, or visual spoilage detection; those are addressed by other systems and datasets in food computing (Kaur et al., 2019, Chen et al., 2024, Kanulla et al., 20 Jan 2026). Its proper domain is the inference of processing level from nutrient composition, optionally augmented by ingredient- or text-derived metadata. Within that domain, its defining contribution is the conversion of a categorical processing ontology into a probabilistic and continuous representation through the FPro score, enabling ambiguity, ranking, and within-class heterogeneity to be represented rather than suppressed (Ispirova et al., 20 May 2025).