Evaluating the Effectiveness of AI-assisted Food Recommendation System on Personalized Nutrition
How to cite
Melo, Grace; Nayga, Rudy; Zahid, Azlan; Wang, Yibin; Yang, Yanjie. 2025. "Evaluating the Effectiveness of AI-assisted Food Recommendation System on Personalized Nutrition." AEA RCT Registry. AEARCTR-0015970. https://doi.org/10.1257/rct.15970-3.1
Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.
* Primary investigator
Abstract
Nutritional needs vary for each person, shaped by their genetic profile, metabolic processes, and microbiome composition. Recent technological advances have stimulated innovative solutions for personalized dietary guidance, with Food Recommendation Systems (FRS) leveraging AI to provide tailored recommendations based on users' preferences and health goals. Yet, the current FRS is limited in integrating health profiles, preferences, and nutrition needs. To address this, the project aims to develop and evaluate the impact of applying an AI and natural language processing (NLP)-driven FRS that personalizes dietary suggestions based on individual health parameters, nutritional goals, and preferences. The impact of this project relies on improving dietary adherence, enhancing user satisfaction, and addressing specific health conditions. Our proposed AI-enabled system will assist users in recommending nutritionally customized foods while also considering suitable options for users with allergies and health concerns. The proposed method utilizes a novel deep generative network architecture to provide personalized recommendations to users based on their health profiles and preferences (precision nutrition). The system will promote healthier eating habits and address diet-related health issues through evidence-based personalized food prescriptions. Our approach is summarized as follows: First, we will utilize two publicly accessible datasets: the nutrition-based food dataset and the person's health profile and preference dataset. The first contains nutritional values (e.g., protein, carbohydrates, fat, fiber, and energy). The second contains information on age, body mass index, basal metabolic rate, targeted energy intake, physical activity level (PAL), and the existence of disease/deficiency (cardiovascular, diabetes, and iron) and food/taste preferences. Second, data will be normalized and fed to the variational autoencoder network, capturing informative features about the users' dietary needs. An NLP algorithm will be trained to adapt to user preferences. A recurrent neural network will generate recommendations to ensure the energy and nutrients align with the user's needs. Loss functions will be deployed to align our system with trustworthy AI principles by ensuring personalized food suggestions closely match expert nutritional guidelines for each user. A user interface (web or app) integrating the recommendation engine will be developed, making it a usable system with end-to-end functionality. Finally, we will implement and evaluate the impact of the developed algorithm on the dietary quality of a representative sample of the US population, including vulnerable groups (healthy living). This step consists of using a randomized controlled trial (RCT) design with a treated group (i.e., those who receive personalized dietary recommendations prescribed by the developed AI system) and a control group (those who receive standard dietary recommendations). Relevance to the IHA Mission: This project aligns with the IHA mission by contributing to key areas such as "Reducing diet-related chronic diseases," "Lowering health care costs," and "Eliminating diet-related health inequities." Through the development and evaluation of personalized dietary recommendations generated by AI, this project directly supports the IHA' s goal of integrating innovative technologies into health-promoting solutions. Further, the project has the potential to bridge the gap between responsible agriculture, healthy living, and public health outcomes, fostering collaborations across disciplines to enhance human health. In the future, we will aim to make prescriptions cost-efficient and, therefore, affordable (responsive agriculture).
Experimental design
We will screen approximately 1,900 U.S. adults through the Forthright Access platform. After an eligibility screener (≥ 18 years; basic English; regular smartphone and internet access; generally healthy adults for whom the Dietary Guidelines for Americans are appropriate — excluding those requiring medically supervised or restrictive diets, e.g., diabetes, celiac/gluten sensitivity, CKD requiring dietary restriction, prescribed or highly restrictive diets, or pregnancy), the research team randomly assigned eligible participants 1:1 to two parallel arms. Accounting for ≈50% eligibility and ≈10% attrition, we anticipate an analytic sample of ≈780 (≈390 per arm).
Both arms use the same mobile/web application, which includes a Track Mode for all participants (log meals; view nutrient and diet-quality summaries) and a Personalized Recommendation Mode for the treatment arm only.
Treatment arm – the app with Track Mode plus the AI/NLP-driven Personalized Recommendation Mode (OpenAI engine constrained to USDA FPED, the Dietary Guidelines, and AMDR ranges; educational only).
Control arm – the same app shell in Track Mode only (tracking + general, non-personalized guidance); no AI-personalized recommendations.
Single-blind from the participant's side; participants are told only that they will receive "nutrition guidance"; the control arm is not informed another group receives AI recommendations. Forthright/the technology team handle randomization; the data-analysis team is masked to arm codes. Dietary intake and behavioral questions are collected at baseline (Week 0) and at the post-intervention endline [registered Week 12; implemented after ~7 weeks under the approved modification]. Intake is measured with ASA24; treatment participants are asked to use the app ≥ once/week and to indicate whether they followed each recommendation; app analytics are logged automatically for engagement/uptake analyses.
Design details
Cited by
No citation data available for this registration yet.