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

The NOVA classification system categorizes foods into four groups based on processing levels: unprocessed or minimally processed (NOVA1), culinary ingredients (NOVA2), processed foods (NOVA3), and ultra-processed foods (NOVA4). While widely used in nutritional epidemiology and public health policy, NOVA's robustness and functionality have been largely unexplored. This study investigates the reliability and consistency of NOVA food assignments among experts. The increasing consumption of industrially processed foods necessitates robust food classification systems to understand the impact on diet quality and health outcomes. Several systems exist, but NOVA is the most prevalent, used extensively in research and policy making, including guiding public health decisions in various countries such as those in Latin America and influencing France's plans to reduce ultra-processed food consumption. Despite its widespread use, a critical evaluation of the system's consistency in application is crucial for its continued validity and effective use in research and policy.

Existing literature indicates that foods are more than the sum of their nutrients, and processing significantly impacts food characteristics. Several food classification systems exist, each using different criteria. Previous studies have highlighted the issue of inconsistent classification within food processing systems, with NOVA's potential for misclassification flagged due to its somewhat vaguely defined categories. However, the extent of this inconsistency and the influence of factors such as detailed ingredient information has not been thoroughly investigated. A systematic review of epidemiological studies using NOVA showed its dominance, emphasizing the need to assess the reliability of this popular system.

French food and nutrition specialists (n=159 for marketed foods, n=177 for generic foods) participated in an online survey. They assigned 120 marketed food products (with ingredient lists) and 111 generic food items (without ingredient lists) to one of the four NOVA groups. Assignment consistency was quantified using Fleiss’ k (range 0–1, with 1 indicating 100% agreement). Hierarchical clustering on principal components (HCPC) identified food clusters based on similar NOVA assignment distributions. Evaluators who did not assess all foods or exhibited atypical assignments were excluded from the analysis. Correspondence analysis (CA) explored NOVA assignment patterns. Cramer's V measured the association between foods and NOVA assignments. Sensitivity analyses were performed by excluding outlier evaluators. The relationship between the most common NOVA assignments (NOVAmaj) and nutritional quality (Nutri-Score, SAIN, LIM, NRF 9.3, energy density) was also examined using chi-squared tests and Kruskal-Wallis tests.

Fleiss’ k values were low (0.32 for marketed foods and 0.34 for generic foods), indicating poor inter-rater reliability. Analysis of marketed foods revealed three clusters: one cluster predominantly assigned to NOVA4 (91% of assignments), and two other clusters showing greater heterogeneity. For generic foods, four clusters emerged: three clusters largely assigned to a single NOVA group (69–79% of assignments), and one cluster (28 foods) with evenly distributed assignments across all four groups. Providing ingredient information did not significantly improve consistency. Many foods, particularly plain unsweetened dairy products and breads, showed highly inconsistent assignments across NOVA groups. Foods classified as NOVA4 (ultra-processed) exhibited a wide range of nutritional qualities, highlighting the disconnect between processing level and nutrient profile. The most common NOVA assignment (NOVAmaj) for each food varied substantially in nutrient profile, indicating that diet quality is influenced by individual food choices within the NOVA4 category rather than the NOVA category itself.

The low inter-rater reliability across different food categories suggests significant ambiguity and inconsistency in the application of the NOVA system. This challenges the current functionality and robustness of the NOVA classification system. The finding that detailed ingredient information didn't significantly improve consistency implies that subjective interpretations and ingrained beliefs about foods influenced the classifications. The observed heterogeneity within the NOVA4 category particularly challenges the established relationship between ultra-processed foods and negative health outcomes, suggesting that nutrient content and consumer choices might be stronger determinants of dietary health than the processing level alone. The study’s findings raise concerns regarding the use of NOVA in epidemiological studies and public health policies, highlighting the need for a revised system that ensures greater consistency and clarity.

The NOVA food classification system demonstrates significant inconsistency in practical application. The low inter-rater reliability and high heterogeneity in assignments, even with ingredient information, highlight limitations in the system's current criteria. Future research should focus on refining the NOVA criteria to enhance clarity and consistency, potentially incorporating a more systematic and hierarchical approach based on objective processing parameters. This is crucial for improving the reliability of epidemiological studies and the effectiveness of public health interventions aimed at reducing ultra-processed food consumption.

The study was conducted with French food and nutrition specialists, limiting generalizability to other populations and cultures. The survey's design, preventing evaluators from revising earlier assignments, might have slightly underestimated consistency. The limited number of food categories in the marketed foods list could have also influenced the results. The reliance on self-reported classifications from experts potentially introduces bias.