Integrating Food Preference Profiling, Behavior Change Strategies, and Machine Learning for Cardiovascular Disease Prevention in a Personalized Nutrition Digital Health Intervention: Conceptual Pipeline Development and Proof-of-Principle Study.

BACKGROUND

Personalized dietary advice needs to consider the individual's health risks as well as specific food preferences, offering healthier options aligned with personal tastes.

OBJECTIVE

This study aimed to develop a digital health intervention (DHI) that provides personalized nutrition recommendations based on individual food preference profiles (FPP), using data from the UK Biobank.

METHODS

Data from 61,229 UK Biobank participants were used to develop a conceptual pipeline for a DHIs. The pipeline included three steps: (1) developing a simplified food preference profiling tool, (2) creating a cardiovascular disease (CVD) prediction model using the subsequent profiles, and (3) selecting intervention features. The CVD prediction model was created using 3 different predictor sets (Framingham set, diet set, and FPP set) across 4 machine learning models: logistic regression, linear discriminant analysis, random forest, and support vector machine. Intervention functions were designed using the Behavior Change Wheel, and behavior change techniques were selected for the DHI features.

RESULTS

The feature selection process identified 14 food items out of 140 that effectively classify FPPs. The food preference profile prediction set, which did not include blood measurements or detailed nutrient intake, demonstrated comparable accuracy (across the 4 models: 0.721-0.725) to the Framingham set (0.724-0.727) and diet set (0.722-0.725). Linear discriminant analysis was chosen as the best-performing model. Four key features of the DHI were identified: food source and portion information, recipes, a dietary recommendation system, and community exchange platforms. The FPP and CVD risk prediction model serve as inputs for the dietary recommendation system. Two levels of personalized nutrition advice were proposed: level 1-based on food portion intake and FPP; and level 2-based on nutrient intake, FPP, and CVD risk probability.

CONCLUSIONS

This study presents proof of principle for a conceptual pipeline for a DHI that empowers users to make informed dietary choices and reduce CVD risk by catering to person-specific needs and preferences. By making healthy eating more accessible and sustainable, the DHI has the potential to significantly impact public health outcomes.