A machine learning algorithm called FoodProX has been developed by researchers, enabling the prediction of the degree of processing in food products.
The foods are scored on a scale of zero (indicating minimal or unprocessed) to 100 (representing highly ultra-processed) by this tool. By bridging gaps in existing nutrient databases, FoodProX provides us with a higher resolution analysis of processed foods.
This development marks a significant advancement for researchers who are investigating the health effects of processed foods.
- FoodProX is a machine learning tool that enables the prediction of the level of processing in a food product.
- The nutritional information from the U.S. Department of Agriculture’s Food and Nutrient Database is utilized by this tool.
- It has been confirmed by the AI tool that over 73% of the U.S. food system is ultra-processed.
The links between “ultra-processed foods” and human health have been under investigation by Northeastern researchers as part of the university-sponsored Foodome project.
As a result of this effort, a machine learning algorithm has been developed by the Center for Complex Network Research, which is said to accurately predict the degree of processing in food products comprising the U.S. food supply.
The findings of the researchers were published in April in Nature Communications.
The machine learning classifier, named FoodProX, utilizes nutritional labeling information provided by the U.S. Department of Agriculture’s Food and Nutrient Database for Dietary Studies as inputs for scoring the level of processing in a given food product.
The algorithm functions by producing an output that represents the likelihood of a respective food falling into one of the four categories that are part of the NOVA food classification system—a system that has been developed by researchers at the University of São Paulo, Brazil, and is widely used in epidemiological studies.
The tool can be accessed by users through the website of the TrueFood research project. The tool allows us to search for a food and view its food processing score. Each product is assigned a single score between zero (indicating “minimally or unprocessed” food) and 100 (representing highly ultra-processed food) by the algorithm.
With the use of FoodProX, gaps in the Nutrient Database for Dietary Studies can be bridged, and “complex recipes and mixed foods and meals” can be classified. This provides us with a higher resolution perspective for examining processed foods.
Consequently, it is observed by the researchers that FoodProX offers us a clearer understanding of the actual level of processing in foods, which is a crucial step for us in studying the health impacts of these foods.
The researchers emphasize how the NOVA system, which categorizes foods into four classifications ranging from “unprocessed or minimally processed” to ultra-processed, is fundamentally limiting because it fails to consider the various degrees of processing within each separate category.
The limitations imposed by the perceived homogeneity of NOVA 4 foods on scientific research and practical consumer guidance regarding the health effects of varying degrees of processing are highlighted by the researchers.
Additionally, the incentives for the industry to reformulate foods towards less processed options are reduced, as investments shift from the ultra-processed NOVA 4 foods to the less processed NOVA 1 and NOVA 3 categories.
In the paper, it is asserted that we believe the nutritional information, which includes the chemicals measured as nutrients in the nutritional facts, somehow encodes the fingerprint of food processing. This is due to the fact that when a food undergoes processing and certain staple ingredients are modified, its chemistry undergoes numerous changes.
The “fingerprinting” method allows researchers to gain insights into the extent of chemical alterations made to a specific food.
According to Menichetti, the lead author of the research and a senior research scientist at Northeastern’s Network Science Institute, we may not know all the chemical fingerprints associated with each process on a one-to-one basis. In fact, the various ways in which a food can be processed cannot even be enumerated.
Ultimately, the team’s AI tool confirmed their previous finding that over 73% of the U.S. food system is ultra-processed, while providing a previously unattainable level of detail. Menichetti mentions that her team is the first to successfully develop an AI tool capable of reliably assessing the chemical content of food.
“It’s the first paper in the space of nutrition and public health where machine learning is leveraged to score foods systematically and reproducibly according to their degree of food processing,” she states.
The team’s work holds significant importance because, as Menichetti points out, there wasn’t much of a data culture in the field of nutrition and health science concerning food processing. This lack of a systematic approach led to less scientifically rigorous discussions about the definition of processing itself.
“When a food cannot be systematically examined and its properties assessed, it becomes challenging to conduct large-scale studies that are comparable across different regions of the world,” Menichetti explains.
“FPro enables us to evaluate an individual’s diet quality and provides predictive power over 200+ health variables,” says Albert-László Barabási, the Robert Gray Dodge Professor of Network Science at Northeastern and co-author of the study.
“It informs us about the effects of replacing processed foods with less processed alternatives of the same item, resulting in personalized dietary adjustments with minimal effort.”
Abstract
Machine learning prediction of the degree of food processing
Despite the accumulating evidence that increased consumption of ultra-processed food has adverse health implications, it remains difficult to decide what constitutes processed food.
Indeed, the current processing-based classification of food has limited coverage and does not differentiate between degrees of processing, hindering consumer choices and slowing research on the health implications of processed food.
Here we introduce a machine learning algorithm that accurately predicts the degree of processing for any food, indicating that over 73% of the US food supply is ultra-processed.
We show that the increased reliance of an individual’s diet on ultra-processed food correlates with higher risk of metabolic syndrome, diabetes, angina, elevated blood pressure and biological age, and reduces the bio-availability of vitamins.
Finally, we find that replacing foods with less processed alternatives can significantly reduce the health implications of ultra-processed food, suggesting that access to information on the degree of processing, currently unavailable to consumers, could improve population health.
“Machine learning prediction of the degree of food processing” by Giulia Menichetti et al. Nature Communications
