Unhealthy diets are a leading cause of non-communicable diseases globally. Traditional dietary guidelines, focusing on food categories, are increasingly being supplemented by considerations of food processing. Studies show that diets rich in unprocessed foods offer greater health benefits compared to those heavy in processed foods. Existing food processing classification systems, such as NOVA, have limitations. NOVA, while widely used, is qualitative, relies on expert judgment, and faces inconsistencies in classification, particularly for composite foods and meals. These limitations hinder research and informed consumer choices. This research addresses these limitations by developing a machine learning approach that objectively quantifies the degree of food processing using readily available nutritional information.

The paper reviews existing literature on the health impacts of processed foods and the limitations of current classification systems. It highlights the need for an objective, reproducible, and scalable method to classify foods based on their processing level. The NOVA classification system, while influential, is criticized for its qualitative nature, inconsistencies, and limitations in classifying composite foods and meals. The authors argue that a more data-driven approach is necessary to overcome these limitations and provide more accurate and reliable assessments of food processing.

The researchers developed FoodProX, a machine learning classifier, using nutrient concentrations as input. The rationale for using nutrients is threefold: (1) nutrient information is consistently reported worldwide; (2) nutrient levels in unprocessed foods are naturally constrained; (3) processing systematically alters nutrient concentrations. The FoodProX algorithm, a multi-class random forest classifier, was trained on data from the Food and Nutrient Database for Dietary Studies (FNDDS) 2009-2010, using foods with existing NOVA classifications. The model outputs a probability vector representing the likelihood of a food belonging to each of the four NOVA categories. A continuous food processing score (FPro) was then derived from this probability vector, ranging from 0 (unprocessed) to 1 (ultra-processed). The algorithm's performance was evaluated using metrics like AUC (Area Under the Receiver Operating Characteristic curve). The robustness of FPro was tested by analyzing its stability against variations in nutrient content across different FNDDS cycles and by incorporating data on food additives from Open Food Facts. Finally, the individual food processing score (iFPro) was calculated for individuals in the NHANES 1999-2006 dataset, weighing food items by their caloric contribution to the diet. This allowed for an assessment of the correlation between iFPro and various health outcomes, using an Environment-Wide Association Study (EWAS). A food substitution analysis was conducted to explore the potential health benefits of replacing highly processed foods with less processed alternatives.

FoodProX demonstrated high accuracy in replicating the manual NOVA classifications (high AUC values across all NOVA categories). The analysis revealed that 73.35% of the US food supply is classified as ultra-processed using FoodProX. The continuous FPro score effectively captured the degree of processing across a wide range of foods and preparation methods. The iFPro score, calculated for individuals in the NHANES dataset, showed significant positive associations with several negative health outcomes, including metabolic syndrome, diabetes, cardiovascular risk factors, inflammation, vitamin deficiencies, and shorter telomere length. Conversely, it showed inverse correlations with levels of various vitamins and phytoestrogens. Positive associations were found with several food additives and contaminants. A food substitution analysis demonstrated that replacing even a single highly processed item with a less processed alternative could significantly reduce iFPro and improve various health markers. Replacing ten items led to even more substantial improvements.

The findings of this study provide strong evidence for the significant impact of food processing on health. FoodProX offers a robust and scalable method for objectively assessing the degree of processing, addressing limitations of existing qualitative methods. The strong correlations observed between iFPro and various health outcomes reinforce the importance of reducing ultra-processed food consumption. The food substitution analysis highlights the potential for effective interventions with minimal dietary changes. The availability of FPro scores, easily accessible through nutritional information already available to consumers, offers a powerful tool for promoting healthier dietary choices and informing public health policies.

This research introduces FoodProX and FPro, providing a novel, quantitative, and easily accessible method for assessing the degree of food processing. The study's findings underscore the significant link between ultra-processed food consumption and adverse health outcomes. The demonstrated effectiveness of simple food substitutions in mitigating these risks emphasizes the importance of making this information readily available to consumers and policymakers. Future research could focus on refining FPro by incorporating information on food additives and expanding its application to various global food systems.

The study relies on self-reported dietary data from NHANES, which might be subject to recall bias and inaccuracies. The analysis focuses primarily on the US population, and the generalizability to other populations with different dietary habits and food environments needs further investigation. The current version of FoodProX primarily uses nutritional information, and the incorporation of data on food additives and processing byproducts could potentially improve its predictive power.