Maintaining proper nutrition is crucial for preserving health and preventing disease. However, what constitutes proper nutrition may vary among individuals; evidence indicates that the effects of diet and even single nutrients can differ considerably because of personal characteristics. This personal variability can be observed through blood markers, such as concentrations of plasma cholesterol and insulin, and captured using a hierarchical multivariate model. We leverage this variability and propose a conditional two-component Bayesian mixture model for generating personalized diet recommendations. The model uses the Nordic Nutrition Recommendations 2023 as a prior for healthy intake and infers individualized recommendations as posterior distributions. The first component identifies dietary options predicted to produce healthy levels across all considered blood markers, while the second selects, among these valid options, the diet closest to predefined personal preferences. The preference component is configurable and, in this study, was used to minimize dietary adjustments to support recommendation adherence while providing well-defined targets for nutrients less critical to concentration regulation. The method was evaluated using nutritional data from two studies: one in prediabetic individuals and one in patients with kidney dysfunction. Numerical simulations showed that the individualized diets could restore or approach normal plasma concentrations when the estimated personal nutrient effects indicated biological feasibility. As the results align with current nutritional literature, the Bayesian approach offers a principled way to leverage observational nutrition data. However, future clinical studies are needed to validate the results and modeling approach before these can be translated into evidence-based personalized nutritional counseling.
Turkia et al. (Sun,) studied this question.