Fig. 3: Health implications and food substitution.
From: Machine learning prediction of the degree of food processing
For each of the 20,047 individuals in NHANES (1999–2006), 18+ years old with dietary records59, we calculated the individual diet processing scores iFProWC. a The average number of unique dishes reported in the dietary interviews, highlighting two individuals \({{{{{\mathcal{A}}}}}}\) and \({{{{{\mathcal{B}}}}}}\), with comparable number of dishes, 12.5 and 13 reported, respectively. b The distribution of average daily caloric intake, showing that individuals \({{{{{\mathcal{B}}}}}}\) and \({{{{{\mathcal{A}}}}}}\) have similar caloric intake of 1894 and 2016 kcal, respectively. c The distribution of iFProWC for NHANES, indicating that individuals \({{{{{\mathcal{A}}}}}}\) and \({{{{{\mathcal{B}}}}}}\) display significant differences in iFProWC, with \({{{{{\mathcal{B}}}}}}\) ’s diet relying on ultra-processed food (iFProWC = 0.9572), and \({{{{{\mathcal{A}}}}}}\) reporting simple recipes (iFProWC = 0.3981) (Figure S13). d We measured the association of various phenotypes with iFProWC, correcting for age, gender, ethnicity, socioeconomic status, BMI, and caloric intake (Section S4). We report the standardized β coefficient, quantifying the effect on each exposure when the Box-Cox transformed dietary scores increase by one standard deviation over the population. For continuous exposures the coefficients are fully standardized, while for logistic regression (disease phenotypes) we opted for partially standardized coefficients to help interpretability (Section S4). Each variable is color-coded according to β, positive associations shown in red, and negative associations in blue. For logistic regressions, p values are associated with two-sided Wald tests, while for multiple linear regressions, p values are determined by two-sided t tests. Here, we show a selection of the 209 variables surviving Benjamini-Hochberg FDR correction with α = 0.05 (*** adj p value < 0.001, ** adj p value < 0.01, * adj p value < 0.05) e Changes in iFProWC when one (orange) or up to ten (yellow) dishes are substituted with their less processed versions, following the prioritization rule defined in Eq. S8. f The impact of substituting different number of dishes on the odds of metabolic syndrome, concentrations of vitamin B12, vitamin C, and bisphenol A, showing that a minimal substitution strategy can significantly alter the health implications of ultra-processed food. Source data are provided in Source Data Figure 3a–f.xlsx.
