NOVA Through Two Lenses: From Nutrients to Food Processing Classification in Finland's National Food Database
This study demonstrates that a machine-learning classifier trained on nutrient composition can effectively transfer across different food databases to identify ultra-processed foods and rank processing intensity, even though assigning all four NOVA groups remains challenging due to inherent definition ambiguities.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
Food is more than just fuel; it is a spectrum of transformation. At one end sit foods in their natural state, like an apple or a raw carrot. At the other end sit products that have been so heavily altered by industry that they bear little resemblance to their original ingredients, such as a sugary breakfast cereal or a packaged snack cake. Between these poles lies a middle ground of foods that have been cooked, preserved, or mixed with simple ingredients like salt or sugar. Scientists have developed a system called NOVA to sort foods into these four categories, hoping to understand how the degree of processing affects human health. The idea is that the more a food is processed, the more likely it is to be linked to poor health outcomes like diabetes or heart disease. However, sorting millions of food items by hand is slow, difficult, and often leads to disagreements even among experts. Furthermore, many national food databases, which track what people eat, do not include these processing labels at all, leaving researchers with a gap in their knowledge.
To bridge this gap, a team of researchers set out to see if a computer could learn to sort foods correctly just by looking at their nutritional content. They focused on a specific machine-learning tool called FoodProX, which was originally trained on American food data to predict NOVA groups based on ten key nutrients, such as sugar, fat, fiber, and sodium. The team wondered if a model trained on European foods could work just as well on American foods, and whether it could successfully label foods in a Finnish national database that had never been sorted before. They gathered a massive collection of over 180,000 food products from a global database, mostly from Europe, and taught the computer to recognize the patterns of nutrients that correspond to each of the four NOVA groups. They then tested this new, European-trained model on unseen European foods and on a separate set of American foods to see if it could generalize its knowledge across borders.
The results showed that the computer was remarkably good at spotting the extremes of the food spectrum. It could reliably identify unprocessed foods and ultra-processed foods with high accuracy. When the researchers asked the model to distinguish between the two most processed categories—the "processed" group and the "ultra-processed" group—it struggled. The model frequently confused these two middle categories, suggesting that the nutritional differences between them are too subtle for a computer to detect with certainty. This confusion was not a flaw in the computer's programming but rather a reflection of the difficulty humans face when trying to draw a clear line between these two groups. Even when the model was applied to the Finnish database, which contained over 4,000 food items with no prior labels, it produced a consistent ranking of foods from least to most processed, placing raw ingredients at the low end and ready-made meals at the high end.
The study revealed that the computer relied heavily on specific nutrients to make its decisions. Sodium, or salt, was the single most important clue, followed closely by total sugars. Foods high in these two nutrients were pushed toward the ultra-processed category, while foods with low fiber and iron were also flagged as likely being highly processed. Interestingly, the model did not rely as much on total energy or saturated fat as one might expect. When the researchers compared their European-trained model with the original American version, they found that while the two models sometimes disagreed on the exact category of a specific food, they largely agreed on the overall order. They both ranked foods in a similar sequence from least to most processed, even if the exact numbers they assigned differed slightly. This suggests that while the precise boundaries between categories are fuzzy, the general trend of increasing processing is clear and measurable through nutrients alone.
The researchers concluded that while nutrient-based models are powerful tools for identifying ultra-processed foods and ranking them by intensity, they cannot perfectly replace human experts in sorting every single food into the four distinct NOVA groups. The difficulty lies in the nature of the classification system itself, where the line between a processed food and an ultra-processed one is often defined by the presence of industrial additives that do not significantly change the nutritional profile. A loaf of bread made with simple ingredients and a loaf made with emulsifiers and preservatives might have nearly identical amounts of sugar, salt, and fat, yet they belong to different categories. Because the computer cannot see the ingredient list or the manufacturing process, it can only guess based on the nutrients it can measure.
Despite these limitations, the study offers a practical path forward. By applying this trained model to the Finnish national food database, the researchers created the first-ever set of processing labels for thousands of Finnish food items. This new data allows scientists to study how food processing in Finland relates to health outcomes without needing to manually label every item. The findings suggest that for large-scale studies, using a computer to flag ultra-processed foods is a reliable strategy, even if the finer distinctions between processed and ultra-processed remain uncertain. The work highlights that while nutrients alone cannot tell the whole story of food processing, they provide a strong signal that can help researchers and policymakers understand the composition of our food supply and its potential impact on public health.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.