(P14-014-24) Hierarchical Ingredient-Level Representation of Dietary Intake Improves Predictive Models of Systemic Inflammation in a Large US Cohort

Objectives: To determine if hierarchical representations of diet improve predictive models of systemic inflammation in a large US cohort.

Methods: The National Health and Nutrition Examination Survey (NHANES) “What We Eat in America” (WWEIA) ranging from 2001-2018 was used in this study. After exclusion for human milk consumption, an average caloric intake < 500 or > 4500 kcal/day, < 18 years of age, active infection of hepatitis or HIV, acute phase response (serum C-reactive protein (CRP) > 10 mg/dL), or missing covariate data, the final sample size was 19,460. Foods reported were disaggregated into ingredients using an automated method developed for this study. Ingredients were used to create a hierarchical food tree annotating leaf nodes with energy-adjusted grams of intake. Hierarchical feature engineering was employed using TaxaHFE to select the optimal set of diet features for machine learning classifiers to predict tertiles of CRP concentration. For each model, 10-fold cross-validation was performed on the training set with 20 percent of the data held out for testing. In parallel, the Dietary Inflammation Index (DII) was calculated and used as a benchmark to compare the performance of these models.

Results: The DII model yielded an accuracy of 0.333, equivalent to random chance. In contrast, the food tree model resulted in an accuracy score of 0.364, a modest improvement. Models that include covariates also demonstrate a slight improvement with the food tree (accuracy 0.521) compared to the DII (0.518). Diet features with the highest importance included “L2_pastas_cooked_cereals,” “L1_fruits,” and “L2_dark_green_leafy_vegetables” each of which were consumed in greater quantity in individuals with low inflammation.

Conclusions: Specific ingredients, selected from a food tree, are stronger predictors of inflammation compared to the DII. Models including covariates demonstrate that person-specific characteristics (e.g. BMI, sex, etc.) are more strongly associated with inflammation than diet, as expected. Overall, our method provides better prediction and resolution (e.g. “pasta/cooked cereals” as opposed to all “grains”) to more precisely inform dietary guidance.

Funding Sources: This work was supported by the United States Department of Agriculture (USDA)/NSF AI Institute for Next Generation Food Systems (AIFS) and the USDA Agricultural Research Service.