Upgraded Estimation of Dietary Intake Using Phenotypic and Biochemical Markers by Supervised Equations: Applicability for Categorizing DQI

Abstract

Objective: Dietary and nutrient intake directly impact health, whereby adherence to certain dietary patterns is linked to positive outcomes. Traditional methods like the Food Frequency Questionnaire (FFQ) and 24-hour recall are subjective, highlighting the need for advanced techniques that incorporate phenotypic and metabolic data. This pilot exploratory study aimed to assess the feasibility of using machine-learning techniques that integrate routinely collected phenotypic and biochemical data to predict adherence to well-characterized dietary quality indices. Method: A total of 138 participants were recruited in the Dietary Deal cross-sectional study to collect data on dietary intake (FFQ, 24-hour recall), biochemical markers, physical activity estimation, quality-of-life questionnaires, and anthropometric determinations. The Mediterranean Diet Adherence Screener (MEDAS 17p), the Alternative Healthy Eating Index (AHEI), the Dietary Approaches to Stop Hypertension (DASH), and a pro-vegetarian model were tested as quality indices. Biochemical and dietary data were integrated using adjusted logistic regressions through STATA (v. 18.0) statistical program to identify biochemical markers associated with food consumption to predict dietary quality. Subsequently, an algorithm based on machine-learning techniques was developed, and the predictive capacity of the obtained models was determined using receiver operating characteristic (ROC) curves and related metrics (area under the curve). Results: A computational algorithm was created for probability classification, adjusted for age, sex, body mass index, physical activity, and SF-36. Key biochemical parameters included glucose, triglycerides, high-density lipoprotein cholesterol, homocysteine, and albumin. Homocysteine (p = 0.007 for AHEI, p = 0.040 for pro-vegetarian), folate (p = 0.039 for DASH, p = 0.019 for pro-vegetarian), and vitamin C (p < 0.001 for AHEI, p = 0.023 for DASH) emerged as significant variables across diet quality indices. The explanatory capacity of the fully adjusted model ranged from R2 = 22.07% to 35.76%, depending on the index. The model’s accuracy ranged from 72.46% to 78.26%, with ROC values between 0.79 and 0.87, indicating moderate to good predictive validity of the training data on itself. Conclusions: This pilot exploratory analysis demonstrates the feasibility of integrating dietary and biochemical data to suitably predict adherence to validated dietary quality indices, Although not intended as a deployable prediction tool, the study provides preliminary evidence supporting the potential of routinely collected clinical data to inform personalized precision dietary advice through objective computational algorithms for precision nutrition implementation.