News

Have scientists discovered a new tool to measure UPFs?

Researchers have developed a new scoring system and database, compiling over 50,000 food items, of which over 1,000 are classified as ultra-processed.

A new research study exploring the prevalence of processed foods in US grocery stores has established a database, collecting over 50,000 food items sold nationwide in supermarkets, including Walmart, Target, and Whole Foods. The new Grocery DB database saw scientists quantify and collate over 1,000 ingredients that they categorised as ultra-processed.

“Highly processed food (HPF) represents a paradigm shift in nutrition science, offering a complex systems perspective on assessing food quality,” Giulia Menichetti, a statistical and computational physicist at the Network Science Institute and Department of Physics at Northeastern University and Harvard Data Science Initiative at Harvard University, told Ingredients Network.

“However, defining HPFs remains controversial, primarily due to the need for deeper fundamental research and better data on food composition and processing to fully map the involved biological mechanisms.”

Ongoing academic discussions, which form a backdrop to the current research, point out that changes in the food matrix characteristic of HPFs may negatively affect nutrient bioavailability, postprandial glycemic responses, and satiety. “While there’s substantial epidemiological evidence linking HPFs to non-communicable diseases, a comprehensive understanding of the underlying mechanisms remains elusive,” Menichetti explained.

As a computational physicist entering the field of nutrition, Menichetti was struck by the lack of large-scale, standardised, open-access repositories – resources that have already transformed fields like genomics and transcriptomics – driving precision medicine forward. “Addressing this gap is critical, especially with the rise of data-hungry tools like large language models,” she said.

GroceryDB was born from this need, aiming to demonstrate how machine learning can unlock insights from real-world, large-scale food composition data. “Our work shows how this information can be translated into actionable metrics, like the degree of food processing, to empower consumer decision-making and inform public health initiatives that strive to improve the overall quality of food environments,” Menichetti added.

“To effectively improve public health, we need a data-driven approach to map and verify these complex biological mechanisms, allowing us to refine and align our definitions and interventions,” she said. With research studies, media coverage and policy reforms bringing UPFs even further to the fore, scientists sought to understand just how present and accessible these foods are to the US population.

“With this spirit, we started our data analysis from grocery stores, the first point of contact with consumers,” she added.

Grocery stores are the ideal starting point for this “temperature check” type because they are a primary touchpoint between consumers and food. “By monitoring and analysing these environments, we can better predict future risks for chronic diseases and design interventions to improve public health at scale,” Menichetti explained.

With GroceryDB, the researchers envision building a database and catalysing a global effort toward open-access, internationally comparable data that advances nutrition security and ensures equitable access to healthier food options.

In the study, researchers relied on nutrients as inputs for several reasons. Firstly, the list of nutrients in food is consistently regulated and reported worldwide. Secondly, their quantities in unprocessed food are constrained by physiological ranges determined by biochemistry. Thirdly, food processing systematically and reproducibly alters nutrient concentrations through combinatorial changes detectable by machine learning.

According to the GS1 UK data crunch analysis, there is, on average, 80% inconsistency in product data. “While the nutrition facts and ingredient list both contribute to food characterisation, the lack of comprehensive, well-regulated global ingredient data led us to prioritise nutritional facts for their portability and reproducibility,” said Menichetti.

Using a lightweight approach, as opposed to a data-hungry model, enables FPro to bridge the gap between curated “model foods” databases used in epidemiology – in which we manually curate NOVA labels and larger nutritional tables but usually do not include ingredient lists – and real-world grocery store products, such as nutrition facts and ingredient lists.

“Given the abundance of what is currently classified as UPFs in our food supply, eliminating this category from our daily lives seems unlikely,” she added. Furthermore, a large population Lancet study indicates that what is currently categorised as UPFs may not be equally bad for our health. Unsurprisingly, there’s confusion.

Researchers identified the FPro score as a unique feature of their work. FPro is a continuous score ranging from zero (representing unprocessed foods like fresh fruits and vegetables and other minimally processed foods) to one (representing UPFs such as instant soups and shelf-stable breads).

FPro goes beyond broad categories and enables ranking foods that are otherwise considered identically processed. “Therefore, we are not particularly fond of global cut-off points since they depend on the food category,” Menichetti explained. However, the researchers have identified that, as a rule of thumb, a value of 0.0-0.4 indicates unprocessed or minimally processed food, 0.4-0.7 suggests processed, and 0.7-1 is a nod to its ultra-processed contents.

“The true value of FPro lies not just in its exact numerical score but in its ability to rank food products within specific categories,” she said. Comparing food types like bread with yoghurt may not be helpful or meaningful. Instead, the actionable insights come from identifying which bread products are less processed than others within the bread category, and the same applies to yoghurt. Consumers can then identify plausible alternatives to the foods they already eat.

The researchers used a machine learning algorithm trained to evaluate the degree of food processing reproducibly, portable and scalable based on nutritional composition data. The algorithm first learns to classify foods into NOVA categories using their nutritional profiles. Then, FPro quantifies the confidence level in distinguishing between NOVA 1 (unprocessed/minimally processed) and NOVA 4 (ultra-processed). These two extreme classes are ranked according to the increasing extent of food processing.

While FPro partially builds on expertise-based classifications like NOVA, the researchers aim to transition FPro into an unsupervised system, independent of manual classifications, by expanding the database to include the “dark matter of nutrition” – chemical signatures of additives and processing by-products.

FPro seeks to progress beyond NOVA, enabling nuanced ranking within specific food categories, such as cheesecakes, facilitating targeted interventions. These interventions can allow for incremental dietary shifts, nudging consumers toward less processed options. “Preliminary calculations suggest that even small changes in dietary trajectories can reduce disease risk, though more foundational research is necessary to deepen our understanding,” Menichetti noted.

The researchers envision tools like GroceryDB playing a pivotal role in global efforts to improve nutrition security and lower diet-related health disparities. “GroceryDB sets the stage for similar initiatives globally – empowering better-informed dietary choices – and highlights the critical importance of open-access, internationally comparable data in advancing global nutrition security,” she added.