An important goal for personalized diet systems is to improve nutritional quality without compromising convenience or affordability. We present an end-to-end framework that converts dietary standards into complete meals with minimal change. Using the What We Eat in America (WWEIA) intake data for 135,491 meals, we identify 34 interpretable meal archetypes that we then use to condition a generative model and a portion predictor to meet USDA nutritional targets. In comparisons within archetypes, generated meals are better at following recommended daily intake (RDI) targets by 47.0%, while remaining compositionally close to real meals. Our results show that by allowing one to three food substitutions, we were able to create meals that were 10% more nutritious, while reducing costs 19-32%, on average. By turning dietary guidelines into realistic, budget-aware meals and simple swaps, this framework can underpin clinical decision support, public-health programs, and consumer apps that deliver scalable, equitable improvements in everyday nutrition.
Diet is one of the most powerful, modifiable drivers of obesity, diabetes, cardiovascular disease, and other non-communicable conditions, yet translating nutrition science into day-to-day meals remains difficult for most people 1 . Personalized diet recommendation systems promise scale and individualization, but many tools still optimize a single goal (taste, calories, or convenience), lack strict standards-based evaluation, and provide limited guidance on how to change as little as possible to eat better 2 . Consequently, there remains a gap between guideline-concordant diets and what recommenders reliably generate in practice 2 . Rule-based and expert-curated systems helped encode guidelines but often sacrificed adaptability and user fit 3 . Subsequent machine-learning approaches -clinical optimization for chronic disease (e.g., DietOS) and IoMT-assisted personalization -improved targeting, yet frequently treated health metrics in isolation and rarely reported controlled benchmarking against USDA nutrient standards 3,4 . Manyobjective/evolutionary methods began balancing adequacy, preferences, and diversity, and clustering/classification pipelines introduced segmentation, but most do not close the loop from what to eat (composition) to how much to eat (portions), a tradeoff which is a key determinant of adequacy, balance, and moderation in real meals 5, 6 and knowledge graph, health-aware recommenders 7 . Generative modeling has accelerated progress towards this goal. Systems, such as Yum-Me, explicitly model both nutrient goals and taste 8 , while recent pipelines leveraging methods like variational autoencoders produce plausible meal plans 9 . More prevalent LLM-based approaches also explore more interactive suggestions 10 . However, systematic reviews highlight inconsistency and factual errors in LLM-generated nutrition, underscoring the need for domain constraints and transparent, multi-objective evaluation 11-14 . For real-world impact in nutrition science and clinical informatics, tools must embed dietary standards, quantify uncertainty, and deliver actionable and minimal-change recommendations that users can implement without overhauling habits 15, 16 . (A) We start from curated meals and train a conditional variational autoencoder (CVAE) that samples realistic food combinations for a chosen archetype (e.g., breakfast) from a structured latent space. A portion assigner initializes standard servings and then adjusts grams to meet USDA RDI/AMDR targets while preserving the combination. Downstream, we evaluate minimal-change substitutions by searching a learned replaceability graph to find swaps that improve nutrition at lower or comparable cost under a portion-based restaurant pricing model. (B) Data and nutrients used for training and evaluation. Counts of meals per time-of-day, the number of foods available for each meal type, and the nutrient panel (4 macronutrients, 11 micronutrients, and 5 diet quality metrics). (C) The encoder/decoder of the CVAE are conditioned on meal type and calorie band via Feature-wise Linear Modulation (FiLM) layers. The decoder outputs food-presence probabilities which feed the portion assignment module.
Here, we present an end-to-end framework for meal generation, and substitution (Figure 1A) designed for public-health impact. We use a training set of 65,202 meals, consisting of 2,019 foods (1,475 of them unique), each associated with 20 nutrients and other nutrient-metrics (Figure 1B), and we partition them into 34 archetypes (e.g. protein and grains) based on food categories and nutritional content. Then we train a Conditional Variational Autoencoder (CVAE), conditioned to these archetypes to generate representative meals that are subsequently compared to real meals (Figure 1C).
Data description. The USDA “What We Eat in America” (WWEIA) component of the National Health and Nutrition Examination Survey (NHANES) was the primary source for all analyses 17- 20 . We used six survey waves (2013-2020), comprising 55,228 respondents and 135,491 meals (Supplementary Table 1). WWEIA provides a hierarchical taxonomy of ingredients, foods, and meals indexed by USDA codes. Foods are constructed from ingredients and meals are constructed from foods. We standardized food codes across survey waves using USDA’s discontinuation and renumbering documentation and retained dropped or revised codes where mapping was unambiguous. The final corpus contained 8,650 food codes and 2,940 ingredient codes. We excluded pre-2013 surveys because discontinuation mappings were incomplete, preventing reliable code harmonization. Definitions for terms used throughout are given in Supplementary Table 2. The pseudocode for the complete framework, which includes the data preprocessing, meal clustering, conditional meal generation, RDI portion assignment, and substitution optimization is provided in the Section 4 of the Supplementary Material. Data Processing. The WWEIA dataset underwent comprehensive p
This content is AI-processed based on open access ArXiv data.