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 (Table A in S1 File). 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 Table B in S1 File. 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 D in S1 File.

Data processing

The WWEIA dataset underwent comprehensive preprocessing to ensure data quality and reduce dimensionality. For surveys with discontinued food code documentation, we created a comprehensive dataset based on the discontinuation code given (Table C in S1 File). For expanded, consolidated, and renumbered food codes, all previous food codes affected were replaced with the most recent food codes. Dropped and revised food codes were kept in the dataset. This yielded a dataset with 120,375 meals and 6,212 foods. Of these meals, 39,749 were breakfast meals, 37,397 were lunch meals, and 43,229 were dinner meals. Data was organized into time-of-day meal-to-food data subsets with corresponding gram amounts. These datasets were further reduced to binary data that represented the presence or absence of food per time-of-day meal. Local Outlier Factorization (LOF) [21] was applied to meals: we used a contamination rate of 0.3%, removing the most extreme composition outliers based on the negative outlier factorization score, which resulted in the removal of 120 breakfast meals, 113 lunch meals, and 130 dinner meals. This 0.3% threshold was chosen after a sensitivity analysis (0.1–1.0%): headline outcomes (median RDI percent reduction, cluster structure) remained stable; see Section A.1.1 in S1 File. To address the sparsity presented in the dataset, i.e., many foods but only several foods per meal, we created a prototype mapping system using a nutrient-aware aggregation algorithm (Section A.1 in S1 File). This approach reduced the food space by 87.5% per subcategory we select prototypes via facility location so that weighted mass coverage ≥ 90% and nutrient fidelity (weighted cosine error or MARE) ≤ 7%, with assignment similarity floor 0.70 (full formulation in Section B in S1 File). Finally, we constructed bootstrap confidence intervals for the mean of each food in each meal-to-food data subset to reduce the list of ingredients to only those ingredients that were representative of the set of foods consumed. The lower bound of the confidence interval was used as a threshold for ingredient removal. This resulted in the filtered data subsets of 528, 627, and 686 foods to 39,435, 37,024, and 42,831 meals for breakfast, lunch, and dinner, respectively.

Clustering analysis

We clustered meals separately by time-of-day (breakfast, lunch, dinner) to preserve meal-specific structure. Each meal was embedded in a hybrid, high-dimensional feature space that combined: (i) nutritional composition (macronutrients, fiber, energy, macro balance/density ratios) and (ii) categorical food composition (gram amounts per WWEIA main and subcategories, e.g., grains, fruits, vegetables, dairy, mixed dishes, snacks/sweets, beverages). Features were z-score standardized within meal type. We then applied an enhanced HDBSCAN [22,23] procedure with meal-type-specific parameters and post-processing cluster merging for stability and coverage. To profile clusters, we contrasted within-cluster vs. complement feature means and controlled multiple testing with Benjamini–Hochberg correction [24] at q ≤ 0.01. We required an absolute mean difference ≥ 0.15 for significance and labeled features “distinctive” at |Δ| ≥ 0.20, reporting Cohen’s d for effect sizes [25]. Detailed feature extraction, clustering parameters, statistical validation procedures, and cluster naming algorithms are described in Section A.2 in S1 File. This process produced stable, interpretable cluster archetypes for each meal type that were used as conditioning variables for generation and evaluation.

Meal generation

Meal generation consisted of two stages: (1) We use a CVAE [26] with a 64-dimensional Gaussian latent z, and a FiLM-conditioned decoder with 3 Dense(512, GELU) blocks driven by concatenated cluster and meal-type embeddings (8 dimensions each). We include a learned pair-specific prior (log-odds of within-pair prevalence) and a hard allowed-foods gate (top-k by prevalence per pair, k = median item count; mask strength 12). Training uses Adam (lr = 5e-4, clipnorm = 0.5) with weighted BCE (dynamic pos-weight), KL (free-bits, annealed β with triangular cycling after warmup), and a count-matching term (full details and hyperparameters in Section B.1 in S1 File). And (2) a simple RDI‑per‑kcal portion assignment that converts presence probabilities to portions for USDA‑aligned reporting while preserving the model’s food combinations.

Portion assignment

Meals are generated in gram units. Portions are initialized from canonical serving sizes (with small jitter) and refined by minimizing signed log₂ deviation from per-meal RDI-per-kcal targets with asymmetric penalties (higher weight for under-consumption of protein, fiber, and key micronutrients; higher weight for over-consumption of sodium, saturated fat, and sugars). Each meal type (breakfast, lunch, dinner) is assigned 25%, 35%, and 40% of a 2,000-kcal daily target, respectively (see Table H in S1 File). This allocation is a conventional three-meal distribution commonly used in dietary pattern and meal-occasion reporting (e.g., WWEIA); we assess robustness to alternative splits in Section B.4 in S1 File. The solution strategy is: (1) initialize portions from canonical serving sizes (with small jitter); (2) run iterative gradient steps to minimize the signed log₂ objective; (3) apply constraints in order: per-item minimum 5 g; group and per-item caps; energy retargeting to hit meal energy; upper-bound-only caps on sodium, sugars, and SFA; beverage grams and beverage energy fraction (≤25% kcal from beverages); soft total-grams cap (900 g) with rebalancing (Section B.2 in S1 File).

Evaluation framework

We conducted a head-to-head evaluation of generated meals across 19,013 meals (6,268 breakfast, 6,393 lunch, and 6,352 dinner meals per significant clusters in Table F in S1 File), assessing established nutritional adequacy and diversity metrics [27–32]. Table A in S1 File shows the metrics used in greater detail. Preprocessing (LOF, nutrient-aware aggregation, and bootstrap-based ingredient filtering) and clustering were fit on the full development set; the CVAE was then evaluated under 5-fold stratified cross-validation on the resulting meal–cluster data. As a leakage-safe check, we re-estimated feature standardization and HDBSCAN clustering within each fold and report test-fold agreement with the baseline clustering in Section A.2.4 in S1 File. For analysis, Cohen’s d was used for effect sizes. Bootstrap resampling (1,000 iterations) generated confidence intervals including the headline RDI alignment (median percent reduction in deviation), cost savings and nutrition gains at substitution operating points, and nutrient-level summaries. FDR correction was applied for multiple comparisons. Model stability was assessed by 5-fold cross-validation.

Meal substitution analysis

For each generated meal, we identify a small set of similar real meals to serve as candidate substitutions. Similarity is based on overall item overlap and composition, while requiring comparable meal energy and a similar number of items. We also consider simple single‑item replacements within the same food category. We define k as the number of items replaced. For each candidate, we compute nutritional improvement as the reduction in average absolute deviation from per‑meal Recommended Daily Intake (RDI) targets (25%/35%/40% of daily RDI for breakfast/lunch/dinner). Cost change is calculated using a portion-based restaurant-style pricing model: for each item we use a grams-per-portion and price-per-portion (from representative U.S. menu and industry portion benchmarks), apply per-item and cross-item caps (e.g., at most one soup bowl per meal), and add a fixed overhead of $2 per meal. Items not matching a named category use a generic side default (150 g, $3 per portion); apparent cost differences between dishes (e.g., some lamb-based vs chicken-and-rice) can therefore be portion- or category-driven rather than reflecting relative retail prices. Full specification and the implementation are given in Section C.1 in S1 File. An alternative grocery per-100g price map (S2 File) is available for sensitivity analyses. We select winning substitutions using a simple trade-off between improvements in nutritional adequacy and cost increases, with optional budget and “no cost increase” constraints. Substitution was attempted for all 19,013 generated meals in the evaluation set; 8,337 (43.8%) had at least one feasible candidate; the proportion with no feasible substitute (56.2%) and the full distribution of signed nutrition and cost changes (including the fraction of selected substitutes with higher cost) are reported in Section C.5 in S1 File. Portions in candidate real-meal matches are taken as observed; for single-item swaps, the grams of the removed item are reassigned to the added item.

Meal clusters capture core U.S. meal archetypes across breakfast, lunch, and dinner

We retained 34 interpretable clusters that map onto common U.S. meal archetypes (i.e., general meal patterns) and differ meaningfully in nutrition (Fig 2). Several clusters reveal behavioral trade-offs. For example a lunch breads & spreads pattern (Fig 2B) pairs breads with a very high fruit share (fruit ratio Cohen’s d = +12.41, q < 10 ⁻ ²⁷¹) and greater macro-diversity (Cohen’s d = +1.97, q < 10 ⁻ ¹¹⁴) but scores lower on overall meal balance (d = −2.05, q < 10 ⁻ ¹⁰³), suggesting that meals within have more snack-like plates rather than balanced entrée plates. Conversely, Mexican entrées concentrate on the main dish and trim the sides (ingredient count Cohen’s d = −0.74, q < 10 ⁻ ⁵²; portion variability d = −0.66, q < 10 ⁻ ³⁷), with strong mixed-dish enrichment (Cohen’s d = +4.47, q < 10 ⁻ ³⁰⁰). Together, these nutrition-anchored contrasts show that the retained clusters for training represent robust meal archetypes spanning energy-dense plates (pizza dinners, sandwich lunches, cereal breakfasts), leaner/fiber-positive options (yogurt across meal types, soups), and snack-style patterns (breads & spreads). Cluster sizes ranged from 210 to 1,871 meals per cluster (median 537), with 12 breakfast, 11 lunch, and 11 dinner clusters (Table F in S1 File). Subsample-and-recluster stability (Adjusted Rand Index vs. baseline) was high for breakfast (ARI median 0.83, 95% CI 0.81-0.84) and moderate for lunch (0.60, 95% CI 0.45-0.66); dinner clusters showed lower ARI (median 0.01; 0.05 when restricted to non-noise assignments), consistent with greater heterogeneity of dinner patterns (Section A.2.4 and Table L in S1 File).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Meal archetypes across time-of-day in embedding space. t-SNE maps of meals from the hybrid feature space (nutrients and WWEIA category grams) show interpretable, compact clusters for breakfast (top), lunch (middle), and dinner (bottom).

Grey points are background meals, and colored islands are clusters that pass false discovery rate (FDR)-controlled distinctiveness tests. Labels reflect the most distinctive food-group and nutrient signals (e.g., Cereal Bowls, Sandwiches, Pizza Meals, Soups & Broths). Donut charts at right summarize each cluster’s food-group composition (WWEIA main categories), illustrating the food-forward basis of the archetypes used for conditional generation and like-for-like evaluation.

https://doi.org/10.1371/journal.pdig.0001367.g002

Generating realistic, uncompromised meals that improve RDI alignment

We compared generated and real meals within each meal‑archetype cluster using nonparametric bootstrap resampling (1,000 iterations per nutrient per cluster), counting an improvement when the 95% bootstrap interval for generated − real lay below zero. Generated meals concentrate in the high‑similarity region while preserving dispersion across clusters (Fig 3A), while nutrition-related characteristics are substantially improving (Fig 3B): adequacy increases across meal types (MAR: + 7.8% breakfast; + 51.3% lunch; + 14.4% dinner); median energy density is higher in generated meals (+104.1%, + 57.1%, + 31.9% for breakfast, lunch, dinner), reported descriptively (Table 1); and micronutrient coverage strengthens (e.g., vitamin C + 26.1%, + 136.2%, + 94.2%). Concomitantly, 97.1% of clusters show lower median deviation with respect to RDI recommendations for generated meals when compared to real same-cluster meals; the median percent reduction in cluster-level medians is 43.2% (95% CI: 11.2–51.6%) for breakfast, 52.1% (95% CI: 37.1–56.6%) for lunch, and 46.0% (95% CI: 32.9–60.9%) for dinner (overall median 47.0%; 95% CI: 11.2–60.9%; Fig 3C). An ablation that applies the same portion optimizer to real-meal food sets decomposes this gain into portion rebalancing and CVAE food-set selection (Table S in S1 File). At the nutrient level (Section B.5 in S1 File), fiber, protein, potassium, and most micronutrients show consistent improvements in median deviation from per-meal targets; by contrast, sodium shows increased deviation for generated meals at lunch and dinner. Total sugars and saturated fat are included in Section B.5 in S1 File, where real meals were already at or below target, percent reduction is not defined and we report deviation levels or mark as not applicable. The largest improvements to RDI score are in the Dinner Cereal Bowls (−80.1% reduction to the RDI deviation), Breakfast Cereal Bowls (−73.0%), and Lunch Deli Sandwich Plates (−67.6%); the only case where deviation was increased was in the Breakfast Pizza Meals (+50.8%). Thus, 33 of 34 clusters showed improved median RDI deviation for generated versus real meals, and the 47% median reduction was not driven by a small subset of archetypes.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Primary evaluation metrics for generated meal assessment. Summary definitions and mathematical forms for the five principal metrics used to evaluate generated meals: Mean Excess Ratio (MER), Mean Adequacy Ratio (MAR), AMDR composite, dietary Diversity (Hill index; the “diversity” axis in Fig 3A is item count, not this metric), and Energy Density. In dietary guidance, lower energy density is often associated with more balanced diets; here we report energy density as a descriptive statistic, generated meals meeting per-meal energy targets may show higher values.

https://doi.org/10.1371/journal.pdig.0001367.t001

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Generated meals improve nutritional targets while preserving variety.

A) Similarity–item count map (axis: number of foods per meal; distinct from the Hill-index diversity in Table 1). Generated meals cluster within their archetypes yet remain widely distributed, indicating preserved variety and realism when matched by archetype. (B) Nutritional shifts by meal type. Relative to matched real meals, MAR (adequacy) increases, AMDR compliance (balance) improves, and energy density increases; selected micronutrients (e.g., vitamin C) and protein adequacy rise. (C) Deviation from USDA targets. Absolute per-meal median RDI deviation drops by 47.0% overall (43.2% breakfast; 52.1% lunch; 46.0% dinner), with significant improvements across most nutrients.

https://doi.org/10.1371/journal.pdig.0001367.g003

To evaluate the performance of our framework against state-of-the-art Large Language Models, we compared 3,400 meals (100 meals per 34 significant clusters) generated by our framework vs. similarly generated by GPT 4o (Section B.3 in S1 File). Results show that in all metrics except diversity, our framework over-performed GPT-4o, which is the most powerful LLMs at the time of the study, likely due to the high complexity and multi-objective nature of the task. More specifically, AMDR compliance was higher for the CVAE when compared to the LLM methods, with 18.9% vs 11.9%, respectively, of meals generated being AMDR compliant. Aggregating macronutrient composition across meals, our framework averaged 12.7% protein, 27.7% fat, and 61.6% carbohydrate, whereas GPT-4o averaged 16.8%, 39.4%, and 43.5%, and as such, GPT-4o’s meals were both high-fat (>35% fat content) and low-carb (<45%). Interestingly, GPT-4o produced more diverse meals (84%) than our framework (78%), which was not one of the objectives for this study. Overall, our framework better satisfies RDI/AMDR-based adequacy and macronutrient balance in the aggregate, while GPT-4o tends to emphasize compositional diversity.

Food substitutions can make meals 10% healthier for 32% less cost

Can we improve existing meals both nutritionally and cost-wise with few ingredient changes? To answer this question, we attempted substitution for all 19,013 generated meals in the evaluation set; 8,337 (43.8%) had at least one feasible candidate and are included in the cost–benefit analysis (3,000 one-hop, 4,146 two-hop, and 1,191 three-hop), drawn from 1,475 unique real foods. The proportion with no feasible substitute (56.2%) and the full distribution of signed nutrition and cost changes, including the fraction where the selected substitute increases cost, are reported in Section C.5 in S1 File. Fig 4A illustrates the per-meal accounting used throughout (nutrient difference and cost) for a specific meal with one and two substitutions (hops). At the knee of the cost–benefit frontier, substitutions deliver meaningful health improvements at lower cost across all hop settings. Nutrition gain is defined as the percent reduction in average absolute deviation from per-meal RDI targets (25%, 35%, and 40% of daily RDI for breakfast, lunch, and dinner). The representative knee points indicate nutrition gains of 5.2% (95% CI: 5.0–5.6%) with 22.0% (95% CI: 20.0–24.5%) savings for 1-hop, 8.1% (95% CI: 7.9–8.4%) with 30.2% (95% CI: 29.5–30.7%) savings for 2-hop, and 10.2% (95% CI: 9.9–10.4%) with 33.8% (95% CI: 32.9–34.9%) savings for 3-hop (Fig 4B). These points summarize the trade-off frontier; the full distribution of outcomes over all 8,337 selected substitutes is given in the Section C.5 in S1 File. Median cost savings and 95% bootstrap confidence intervals are reported in Section C.1 in S1 File and in the cost-sensitivity tables (Tables N and Table O in S1 File). The reported savings are relative to the stated portion-based pricing model; the pricing level is fixed at a single point in time and does not capture geographic or temporal variation. The composition of gains shifts systematically with hop count, with the number of hops correlating with the across-category by intra-category substitution per meal ratio, suggesting that the method must resort to solutions from other categories to maximize benefit when multiple substitutions are allowed (Fig 4C). With one-hop substitutions, improvements are nutrition-led, accompanied by moderate cost relief and small adjustments to items/portions. Allowing two-hop substitutions maintains strong nutrition contribution and raises ease of adoption, indicating many options that stay close to the original plate. Under three-hop, the center of mass pivots toward savings while nutrition’s share moderates; ease remains comparable. In other words, expanding flexibility widens access to deeper discounts while still improving nutrition, whereas tighter constraints emphasize nutrition-first improvements with smaller changes. Trade-off diagnostics are consistent with these patterns. Together, these results show that along the empirical frontier there are two practical operating modes, namely a local, nutrition-forward improvement with meaningful but moderate savings, to a more far-reaching, budget-forward improvements that achieve larger discounts while still moving nutrition upward.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Cost–benefit of machine-guided meal substitutions.

(A) A real meal (top) and two of its generated one-hop (bottom left) and two-hop (bottom right) substitutes. The middle plot shows per-nutrient ratios (substitute/original; log scale) with the vertical line at parity. Green points to the right indicate increased progress toward daily targets; red to the left indicate decreases. Insets show benefits of 2-hop swaps include a reduction in cholesterol and an increase in vitamins. (B) Population-level cost–benefit curves summarizing, for one-, two-, and three-hop substitutions, the trade-off between nutrition gain and cost savings as the policy parameter θ varies. Nutrition gain is the percent reduction in average absolute deviation from per-meal RDI targets (25%, 35%, and 40% of daily RDI for breakfast, lunch, and dinner, respectively). Stars indicate the selected operating points.; error bars or shaded regions show 95% bootstrap CIs (1,000 resamples) for median nutrition gain and cost savings. (C) Category transition heatmaps for each starred operating point. Each cell shows the probability that the row category is swapped for the column category. Warmer colors denote more frequent directed swaps. Marginal summaries indicate net adds/removes, and clustering exposes motifs (e.g., more additions from vegetables/grains and fewer removals from high cost/low value categories), consistent with the cost–benefit gains. (D) Category-level swap scatters at the same operating points. Each point is a category, with x-axis and y-axis denoting the probability for the added and removed category, respectively. Points on the diagonal reflect in-kind replacement; right/down shifts indicate net additions, up/left indicate net removals. With higher hop counts, patterns concentrate in add-heavy, easy-to-apply substitutions, consistent with an improved cost–benefit. Meal images shown in panel A were generated by the authors using ChatGPT/OpenAI and are not reproduced or adapted from third-party published or copyrighted sources.

https://doi.org/10.1371/journal.pdig.0001367.g004