Aug 2026· Journal of NutriLife· pp.
101786
· 0 citations· 37 references
Medicine
TL;DR
LLMs can generate estimates of nutrient content from narrative descriptions of Peruvian meals, but current performance levels fall short of the precision required for clinical deployment or commercial consumer-facing applications.
Abstract
Background
Artificial intelligence applications have been developed to predict the nutrient content of meals. However, none have been evaluated in the context of Peruvian cuisine, characterized by diverse ingredients and recipes. We assessed whether large language models (LLMs) could predict the nutritional content of Peruvian meals.
Methods
Using a dataset of 510 unique lunch images extracted from a Peruvian cookbook, we compared nutrient values from recipe data against predictions generated by LLMs (Gemma-3 4B, 12B, and 27B). The LLMs were given the meal name and a photograph and prompted to produce narrative descriptions of the meal. Using the descriptions, the same LLMs were prompted to estimate six nutrients: energy (kcal/serving), protein (g/serving), carbohydrates (g/serving), iron (mg/serving), vitamin A (μg/serving), and zinc (mg/serving). Agreement proportions and errors metrics were calculated against the values from the recipe book.
Results
The 27B LLM achieved the highest agreement proportions across most nutrients-calories (45%), carbohydrates (31%), iron (15%), vitamin A (19%), and zinc (31%)-while the 12B model performed best for protein (70% agreement). The 27B model yielded the lowest mean absolute error (MAE) for calories (108 kcal), carbohydrates (26 g), iron (4 mg), and zinc (1 mg). The 12B LLM had the lowest MAE for protein (6 g) and vitamin A (667 μg). The 4B LLM showed the poorest performance across metrics.
Conclusions
LLMs can generate estimates of nutrient content from narrative descriptions of Peruvian meals, but current performance levels fall short of the precision required for clinical deployment or commercial consumer-facing applications.
Several studies have evaluated the ability of Large Language Models (LLMs) for meal planning, yielding positive outcomes. These models can process natural language inputs and leverage learned knowledge from their pretraining to generate meal plans. In this work, we investigate the ability of LLMs to analyze the suitability of given recipes for diabetes. The primary challenge for LLMs is to retrieve relevant dietary guidelines for diabetes, decompose recipes into ingredients and cooking methods, and apply these guidelines to determine the recipe's suitability. To study these challenges, we employ three kinds of prompts namely, (i) Direct Query Prompt (ii) Context-Guided Prompt, and (iii) Exemplary Context Prompt that incorporate different levels of diabetes dietary guidelines from medical sources. We introduce a benchmark dataset curated for this investigation consisting of 7607 recipes that include 3807 recipes suitable for diabetes and 3800 recipes not suitable for diabetes. Our results demonstrate that most LLMs are cautious in predicting recipes as suitable to prevent detrimental outcomes. Further, the models that can reason using the dietary guidelines performed better in predicting the suitability of recipes for diabetes. Overall, Mistral-7B and Llama 70B showed superior performance to their counterparts.
Revathy Venkataramanan, Aditya Luthra, Venkatesan Nadimuthu et al.· 0 citations
Nutritional deficiencies remain a significant global public health challenge, while existing dietary assessment and recipe recommendation systems often operate independently, limiting their ability to provide integrated and personalized nutritional guidance. This article presents an integrated artificial intelligence framework developed using Sri Lankan dietary reference intake guidelines, culturally specific food composition data, and a curated corpus of Sri Lankan recipes to support household nutritional guidance. The proposed framework was developed as part of this Sri Lankan study to address local nutritional challenges while providing a methodology that can be adapted to other countries by replacing country-specific dietary reference standards, food composition databases, and recipe repositories with those of other countries. The nutritional model employed adequacy ratio-based features with a random forest classifier, achieving 88.12% accuracy and a macro area under the curve of 0.91. The semantic module used Sentence-Bidirectional Encoder Representations from Transformers embeddings with fuzzy ingredient matching to achieve 86.84% classification accuracy under stratified cross-validation. By linking predicted deficiencies to context-aware recipes, the system transforms analytical insights into actionable meal recommendations. The results demonstrated that the framework achieved stable performance and showed strong potential for practical application.
L. Weerasinghe, Methusala Perera, Shahmi Mohamed et al.· Artificial Intelligence in H...· 0 citations
Computational nutrition needs precise ingredient data, but current databases are incomplete, inconsistent, and built for human reference rather than automated reasoning. LLMs could help fill these gaps, but single-pass outputs are unreliable and can introduce silent errors into downstream computation. We present a quality-controlled LLM pipeline for ingredient data acquisition that combines robust statistical estimation, domain-specific invariant checks, and a web-fetch fallback. An illustrative Heap's Law fit to 233 recipes suggests that unique-ingredient growth is sub-linear and front-loaded: the projected ratio of unique ingredients to recipes falls from 1.74 at 100 recipes to 0.19 at 5,000. For each ingredient attribute, repeated LLM queries are treated as samples from a model-induced answer distribution, and we apply robust point estimators and normalised confidence scores across numerical, Boolean, multiple-choice, open categorical, and optional integer types. An invariant guard layer enforces nutritional and logical self-consistency within each ingredient record. Minor numeric inconsistencies are reconciled via a linear program that minimises worst-case percentage deviation while preserving semantic zeros, and major violations are escalated to web-evidence-grounded repair, then human review only if that fails. On a curated 30-ingredient reference set, the pipeline achieves 98.4% exact match on nutrient flags and cuts median absolute percentage error on nutrient ratios from 31.9% for the median-aggregated baseline to 10.1%, a reduction of 21.8 percentage points, at an API cost of about $1 per ingredient. This frames LLM-assisted database construction as a controlled data-engineering workflow that makes uncertainty operational rather than discarding it.
General-purpose Large Language Models (LLMs) like Llama, GPT, and Mistral struggle with domain-specific challenges in food and nutrition, where data is fragmented, heterogeneous, and semantically complex. While fine-tuned LLMs have shown success in healthcare and life sciences, similar progress in food domains has been limited, largely due to the lack of high-quality, task-specific datasets. We present FoodBench, a curated benchmark dataset of question–answer pairs designed for training and evaluating LLMs in food and nutrition. It spans key tasks such as nutrient estimation, food traffic-light classification, synonym linking, cooking measurement conversion, and food named-entity recognition and linking. FoodBench enables robust performance evaluation across zero-, one-, and few-shot settings, laying the groundwork for trustworthy, domain-adapted language models. This resource supports advances in personalized nutrition, dietary assessment, and food system innovation. Evaluation of four general-purpose LLMs (Llama 3, Mistral, Gemma, Gemini) on FoodBench tasks shows limited performance across nutrient estimation, traffic-light classification, and food interoperability, even with few-shot prompting. These results highlight the need for domain-specialized LLMs fine-tuned on food data, while establishing FoodBench as a benchmark not only for assessing general-purpose models but also for guiding and evaluating fine-tuning efforts.
T. Eftimov, Ana Gjorgjevikj, Matej Martinc et al.· Scientific Data· 0 citations
BACKGROUND
Personalized meal planning by registered dietitian nutritionists (RDNs) is time-intensive. Large language models (LLMs) may automate drafting meal plans, but their nutritional accuracy in clinical practice is uncertain.
METHODS
In this proof-of-concept study, five outpatient RDNs and four LLMs (Gemini, CoPilot, ChatGPT 4.0, and customized ChatGPT 4.0) each generated 3-day meal plans for five validated clinical scenarios. Effectiveness was defined as accuracy in meeting pre-specified energy, protein, carbohydrate, fat, and sodium targets. Time to create plans and RDN comfort (self-rated confidence in nutritional accuracy and clinical appropriateness on 1-5 Likert scale) were recorded. Three independent RDNs, blinded to source, analyzed nutrient content using Nutritionist Pro. Group differences were assessed with t-test and ANOVA.
RESULTS
All LLMs and RDNs produced feasible meal plans. LLMs generated meal plans in under 1 min, whereas RDNs required a mean of 44 min per scenario. RDNs reported comfort levels ranging from 3.8 to 4.8. Across most scenarios, LLM plans delivered a smaller proportion of requested energy than RDN plans, which more consistently approached energy targets. Both groups performed similarly for the Mediterranean diet scenario. Overall, protein accuracy did not differ. However, in chronic kidney disease, LLMs undershot the guideline-based protein target, while RDNs tended to modestly exceed it. Accuracy for low-carbohydrate, fat, and sodium diets was comparable.
CONCLUSION
LLMs can rapidly generate clinically plausible meal plans but are less reliable than RDNs in achieving prescribed energy and selected macronutrient goals. Prompt precision is essential for nutrient-specific targets. A hybrid model in which RDNs refine LLM-generated drafts may leverage efficiency without sacrificing clinical accuracy.
M. Mundi, Osman Mohamed Elfadil, Danielle P. Johnson et al.· Nutrition in clinical practi...· 0 citations
Personalized nutrition recommendations encounter difficulties in aligning users′ dietary preferences with appropriate foods in extensive nutritional databases comprising thousands of products. This research introduces a multinutrient clustering framework that examines 8790 foods from the USDA National Nutrient Database for Standard Reference, Release 28, utilizing 23 nutritional attributes, including macronutrients, vitamins, and minerals. We thoroughly analyze K‐means and agglomerative clustering algorithms across various configurations (k = 2 − 8), finding that K‐means with eight clusters yields optimal performance, achieving a silhouette score of 0.273 and semantically interpretable dietary categories. The suggested technique shows a 46.4% improvement over suggestions based on popularity and a 273.0% improvement over recommendations based on a single nutrient. This was shown by a thorough evaluation utilizing Precision@K, NDCG, and fivefold cross‐validation (mean silhouette 0.265 ± 0.006). Users can set their own dietary preferences using nutrient sliders, six preset configurations (high protein, low carb, high fiber, low sodium, high calcium, and low calorie), and three configurable weighting strategies (equal, prioritized 3×, and focus‐only). The system then gives them real‐time recommendations (in less than 2 s) with clear similarity scores. A systematic examination with 50 automated test questions shows that the system works well in a wide range of dietary situations. The huge effect sizes (46.4% and 273.0% improvements) suggest that the results are statistically significant. This research connects computational nutrition studies with real‐world dietary advice. It offers an open‐source, understandable system for evidence‐based meal planning that fills important holes in current prediction and classification methods by allowing personalized, multinutrient food suggestions with clear reasoning behind the choices.
Rajkumar Sarker, Kazi Farhan Hasan Tanjim· International journal of foo...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
