Ultra-processed foods: how functional is the NOVA system?

The study examines whether the NOVA food classification system, which groups foods by the nature, extent, and purpose of processing, yields consistent assignments when used by specialists. With growing use of NOVA in nutritional epidemiology and public policy (e.g., dietary guidelines and initiatives to reduce ultra-processed food intake), assessing its robustness and functionality is critical. The authors highlight the complexity of modern food processing and the prevalence of NOVA in research (used in 95% of studies on processing and health outcomes between 2015 and 2019), motivating an evaluation of user consistency in assigning foods to NOVA groups.

Prior research indicates multiple systems classify foods by processing using varied criteria. NOVA is the most commonly used. A previous study reported lower inter-rater reliability for NOVA than for two other processing classification systems, suggesting NOVA’s groups may be insufficiently defined. Policy adoption of NOVA (e.g., in Latin America and France) underscores the importance of reliable classification. The paper also references discussions that processed food classifications are conceptually complex, mixing technological and socio-cultural aspects, and that ultra-processed food intake is associated with certain nutrient intake patterns across countries.

Design: Cross-sectional online survey of French food and nutrition specialists who assigned foods to NOVA groups (NOVA1–NOVA4). Materials: Two food lists were used. (1) Marketed foods: 120 commercial products with ingredient information. (2) Generic foods: 111 common food items without ingredient information. Participants: Initially, 196 evaluators assessed marketed foods and 202 assessed generic foods; 144 completed both lists. Exclusions: Evaluators were excluded if they did not assess all foods on a list (30 marketed; 24 generic) or failed the quality control test (7 marketed; 1 generic), yielding 159 evaluators for marketed foods and 177 for generic foods. Total assignments: 19,080 (marketed) and 19,647 (generic). Quality control: Embedded test foods with expected assignments; evaluators producing erroneous assignments for more than one test food were excluded. Evaluators also reported confidence (very low to high) for each assignment. Data description: For each food, percentages of assignments to each NOVA group were computed. Counts of foods assigned to 1, 2, 3, or 4 groups were tallied. Analytical methods:

• Correspondence analysis (CA) on frequency tables of assignments per list to visualize assignment patterns; association quantified using Cramer's V (0–1; higher indicates more consistent assignments).
• Inter-rater agreement quantified using Fleiss’ κ, with means from 1000 bootstrapped samples (alpha 5%), stratified by professional background. Each bootstrap sample included at least 10 evaluators per background (sample size 70) to assess whether similar expertise improved concordance.
• Clustering: Hierarchical clustering on principal components (HCPC) using Ward’s method to identify clusters of foods with similar assignment distributions; number of clusters chosen to minimize within-cluster variation. Percentage of assignments to NOVA1–NOVA4 computed per cluster.
• Sensitivity analyses: Outlier evaluators detected in contingency tables per Lindskou et al.; analyses repeated excluding outliers to test robustness of κ and cluster structures.
• Nutritional quality analyses: For each food, defined its most common assignment (NOVAmaj). Explored relationships between NOVAmaj categories and nutrient profile metrics (Nutri-Score classes; SAIN,LIM classes) using chi-squared tests; compared distributions of Nutri-Score, SAIN, LIM, NRF 9.3, and energy density across NOVAmaj categories using Kruskal–Wallis tests; boxplots prepared. Nutrient data from CIQUAL; NRF 9.3 per standard methodology. Software: R 4.0.2; packages: irr (κ), FactoMineR (CA, HCPC), DeskTool (Cramer's V). Significance at alpha 0.05.

• Assignment distributions: Marketed foods were assigned predominantly to NOVA4 (80.0% of the 19,080 assignments), followed by NOVA3, with very few NOVA1 or NOVA2 assignments. Generic foods were most frequently assigned to NOVA4 (45.3% of 19,647), then NOVA3, NOVA1, and NOVA2.
• Evaluator confidence: Predominantly high or intermediate; fewer than 10% of assignments had low or very low confidence for either list.
• Agreement levels: Fleiss’ κ was low—0.32 (marketed; 159 evaluators) and 0.34 (generic; 177 evaluators). Professional background did not materially affect agreement (e.g., κ ranges ~0.28–0.37 across groups). Removing outlier evaluators increased κ by at most 0.03.
• Association strength per food: Cramer's V indicated notable heterogeneity (0.58 for marketed; 0.59 for generic).
• Unanimity: Only 3 marketed foods and 1 generic food were assigned to the same NOVA group by all evaluators. Most foods were placed into 2–4 different NOVA groups.
• Clusters of assignment patterns: • Marketed foods (N=120): Three clusters.
  • T (90 foods; 14,310 assignments): Highly homogeneous; 90.7% NOVA4, 8.6% NOVA3, 0.7% NOVA2, 0.1% NOVA1.
  • U (24 foods; 3,816 assignments): Mixed; 53.6% NOVA4, 40.3% NOVA3, 4.8% NOVA2, 1.4% NOVA1.
  • V (6 foods; 954 assignments): Diverse; 48.0% NOVA3, 23.3% NOVA1, 14.5% NOVA2, 14.3% NOVA4. • Generic foods (N=111): Four clusters.
  • W (65 foods; 11,505 assignments): Mostly NOVA4 (69.5%); NOVA3 25.8%; NOVA2 3.7%; NOVA1 1.0%.
  • X (28 foods; 4,956 assignments): Heterogeneous across all groups; NOVA3 53.2%, NOVA4 16.8%, NOVA1 16.6%, NOVA2 13.4%.
  • Y (5 foods; 885 assignments): Predominantly NOVA2 (74.5%); NOVA1 13.3%; NOVA3 10.1%; NOVA4 2.2%.
  • Z (13 foods; 2,301 assignments): Predominantly NOVA1 (78.9%); NOVA2 11.1%; NOVA3 7.8%; NOVA4 2.3%.
• Examples of ambiguity: Plain unsweetened dairy products and breads frequently spanned multiple NOVA groups; commercial orange juice often assigned to NOVA4 versus fresh juice mostly NOVA1; coffee generally assigned NOVA1 despite industrial roasting, reflecting cultural perceptions permitted by NOVA’s rules.
• Nutritional quality: Foods most commonly assigned as NOVA4 (NOVA4maj) spanned all nutrient profile classes. For marketed foods, 26% of NOVA4maj were Nutri-Score A and 35% were SAIN,LIM class 1. For generic foods, 18% of NOVA4maj were Nutri-Score A and 32% SAIN,LIM class 1. Marketed NOVA3maj foods showed better Nutri-Score, SAIN, LIM than NOVA4maj, with no difference in energy density or NRF 9.3. For generic foods, NOVA1maj > NOVA3maj > NOVA4maj in nutritional quality; NOVA2maj consistently worst.

The study demonstrates low inter-rater consistency when applying NOVA, independent of professional background and unaffected by providing ingredient lists. High self-reported confidence suggests evaluators relied on subjective or contextual interpretations, reflecting ambiguities within NOVA’s criteria. Contradictions (e.g., yogurt cited as NOVA1 example but involving fermentation characteristic of NOVA3 and ingredients sometimes linked to NOVA4) and non-hierarchical, mixed technological/formulation rules promote divergent decisions. Heuristics like number of ingredients or perceived industrial origin likely prompted NOVA4 assignments even without NOVA4-characteristic additives, while traditional processes (e.g., coffee roasting) permitted NOVA1 classifications despite industrial processing. The heterogeneous clusters and widespread multi-group assignments indicate that current NOVA criteria do not support systematic, reproducible classification. Nutritional profiling results show that NOVA4maj foods include items with favorable nutrient profiles, suggesting that labeling based solely on NOVA may conflict with front-of-pack nutrient labels and potentially confuse consumers. The findings call into question the reliability of epidemiological associations and policy guidance that depend on NOVA classifications without addressing its operational ambiguity.

NOVA assignments by trained specialists showed substantial inconsistency, with low Fleiss’ κ and broad cross-group variability, regardless of ingredient information. The results indicate that NOVA, as currently specified, lacks the robustness and functionality needed for reliable food classification. The authors recommend improving NOVA through clearer, hierarchical decision rules and disentangling processing intensity (thermo-mechanical energy/unit operations) from formulation (ingredients, additives). Developing a standardized processing-intensity indicator and a decision tree could enhance reproducibility and help clarify whether observed health associations relate more to processing structure or to composition. Interdisciplinary collaboration among food process engineering, food science, nutrition, and epidemiology is suggested to refine classification frameworks and better inform research and policy.

• Marketed foods covered only three categories; findings may differ with broader product categories.
• All evaluators were French; although experts in nutrition/food technology, cultural influences cannot be fully excluded.
• Survey design did not allow revising earlier assignments or real-time discussion, which might have marginally improved consistency.
• Representativeness of evaluators may be limited.