**Background:** Unhealthy diet is a major risk factor for non-communicable diseases, accounting for 70% of mortality worldwide. While food processing has been linked to adverse health outcomes, existing classification systems like NOVA are qualitative, cover only ~35% of foods in the USDA database, and treat all ultra-processed foods as a homogeneous category. This limits both research and consumer guidance. The authors hypothesized that nutrient profiles—which are consistently regulated and reported—systematically change with processing and can be used to predict the degree of processing via machine learning.
**Methods:** The authors developed FoodProX, a multi-class random forest classifier trained on nutrient concentrations from FNDDS 2009–2010 for 2,484 foods with manual NOVA labels. The model outputs a probability vector {p1, p2, p3, p4} for NOVA classes 1–4. From this, they derived FPro, a continuous processing score ranging from 0 (unprocessed) to 1 (ultra-processed), defined as FPro_k = (1 - p1^k + p4^k)/2. The individual-level score iFPro was calculated for 20,047 NHANES participants (1999–2006) as the calorie-weighted average of FPro across all consumed foods. An Environment-Wide Association Study (EWAS) was conducted using survey-weighted generalized linear models adjusted for age, sex, ethnicity, BMI, caloric intake, and socioeconomic status, with Benjamini-Hochberg FDR correction (α = 0.05).
**Key Results:** FoodProX achieved high predictive performance across all NOVA classes: AUC of 0.9804 ± 0.0012 for NOVA 1, 0.9632 ± 0.0024 for NOVA 2, 0.9696 ± 0.0018 for NOVA 3, and 0.9789 ± 0.0015 for NOVA 4. Applying FoodProX to all foods in FNDDS revealed that 73.35% of the US food supply is ultra-processed (NOVA 4), 18.36% processed (NOVA 3), 7.39% unprocessed (NOVA 1), and 0.90% processed culinary ingredients (NOVA 2). The median iFPro_WC for the US population was 0.7872. After FDR correction, 209 variables showed significant associations with iFPro. Higher iFPro was positively associated with metabolic syndrome, diabetes (fasting glucose), Framingham and ACC/AHA risk scores, blood pressure, trunk fat, subscapular skinfold, blood insulin, triglycerides, C-reactive protein, homocysteine, and methylmalonic acid. Inverse associations were found with HDL cholesterol, telomere length (indicating higher biological age), and blood levels of vitamin B12 and vitamin C. Novel findings included positive associations with acrylamide, polycyclic aromatic hydrocarbons, benzenes, furans, PCBs, perfluorooctanoic acids, phthalates, and bisphenol A. The substitution analysis showed that replacing a single food item with a less processed alternative reduced median iFPro_WC by 12.15% (from 0.7872 to 0.6915), corresponding to a 12.25% decrease in odds of metabolic syndrome, 8.47% lower urinary bisphenol A, and increases of 4.83% in vitamin B12 and 12.31% in vitamin C. Substituting 10 items led to a 37.03% decrease in iFPro_WC.
**Clinical Implications:** This study provides a reproducible, scalable tool for quantifying food processing that goes beyond the binary or four-category NOVA classification. The continuous FPro score reveals substantial heterogeneity within the ultra-processed category, enabling more nuanced research into health effects and practical dietary guidance. The finding that even single-food substitutions can meaningfully improve health outcomes suggests that providing consumers with processing information could enable effective dietary interventions. The authors note that FPro currently requires nutrient data and is best suited for ranking foods within categories, but could be extended with additive and processing byproduct data for fully unsupervised classification. This approach could help address the shift from food security to nutrition security and support the UN Sustainable Development Goals.