大语言模型在食品、食品营养及药食同源方面

Understanding the core principles of nutrition is essential in the contemporary context of abundant and often contradictory dietary advice, to empower individuals to make informed dietary choices and manage diet-related non-communicable diseases. The role of Artificial Intelligence (AI) in providing nutritional information is increasingly prominent, but its reliability in this domain is not well-established yet. This study compares the nutrition knowledge of state-of-the-art Large Language Model (LLM)-based conversational agents and chatbots with that of human subjects having different levels of nutrition knowledge. The “General Nutrition Knowledge Questionnaire–Revised” (GNKQ-R) was administered to four LLMs (ChatGPT-3.5, ChatGPT-4, Google Bard, currently known as Google Gemini, and Microsoft Copilot), using zero-shot prompts. Responses were scored in accordance with the GNKQ-R’s guidelines. The average performance of AI systems across all LLMs was 77.3 ± 5.1 out of 88, comparable to that of dietetics students and significantly higher than English students. ChatGPT-4 scored highest among the LLMs (82/88), surpassing both groups of students (dietetics: 79.3/88, English: 67.7/88) as well as all other demographic groups. In “Dietary Recommendations”, ChatGPT-3.5 and ChatGPT-4 demonstrated comparable performance to dietetics students. ChatGPT-4 excelled in “Food Groups”, outperforming all human groups. In “Healthy Food Choices”, ChatGPT-4 achieved a perfect score, indicating a deep understanding of this subject. ChatGPT-3.5 excelled in “Diet, Disease and Weight Management”. Variations in the performances of the LLMs across different sections were observed, suggesting knowledge gaps in certain areas. Some of the tested LLMs, particularly ChatGPT-3.5 and ChatGPT-4, showed proficiency in nutrition knowledge, rivaling or even surpassing dietetics students in certain sections. This indicates their potential utility in nutritional guidance. However, this study also identified nuances and specific details where LLMs lack compared to specialized human education. The study highlights the potential of AI in public health and educational settings. However, LLMs may be limited in their capacity to generate personalized dietary advice that accounts for clinical complexity and individual variability, reinforcing the indispensable role of expert human judgment.

ABSTRACT Large language models (LLMs) such as ChatGPT are increasingly positioned to be integrated into various aspects of daily life, with promising applications in healthcare, including personalized nutritional guidance for patients with chronic kidney disease (CKD). However, for LLM-powered nutrition support tools to reach their full potential, active collaboration of healthcare professionals, patients, caregivers and LLM experts is crucial. We conducted a comprehensive review of the literature on the use of LLMs as tools to enhance nutrition recommendations for patients with CKD, curated by our expertise in the field. Additionally, we considered relevant findings from adjacent fields, including diabetes and obesity management. Currently, the application of LLMs for CKD-specific nutrition support remains limited and has room for improvement. Although LLMs can generate recipe ideas, their nutritional analyses often underestimate critical food components such as electrolytes and calories. Anticipated advancements in LLMs and other generative artificial intelligence (AI) technologies are expected to enhance these capabilities, potentially enabling accurate nutritional analysis, the generation of visual aids for cooking and identification of kidney-healthy options in restaurants. While LLM-based nutritional support for patients with CKD is still in its early stages, rapid advancements are expected in the near future. Engagement from the CKD community, including healthcare professionals, patients and caregivers, will be essential to harness AI-driven improvements in nutritional care with a balanced perspective that is both critical and optimistic.

The integration of large language models (LLMs) into clinical nutrition marks a transformative advancement, offering promising solutions for enhancing patient care, personalizing dietary recommendations, and supporting evidence-based clinical decision-making. Trained on extensive text corpora and powered by transformer-based architectures, LLMs demonstrate remarkable capabilities in natural language understanding and generation. This review provides an overview of their current and potential applications in clinical nutrition, focusing on key technologies including prompt engineering, fine-tuning, retrieval-augmented generation, and multimodal integration. These enhancements increase domain relevance, factual accuracy, and contextual responsiveness, enabling LLMs to deliver more reliable outputs in nutrition-related tasks. Recent studies have shown LLMs’ utility in dietary planning, nutritional education, obesity management, and malnutrition risk assessment. Despite these advances, challenges remain. Limitations in reasoning, factual accuracy, and domain specificity, along with risks of bias and hallucination, underscore the need for rigorous validation and human oversight. Furthermore, ethical considerations, environmental costs, and infrastructural integration must be addressed before widespread adoption. Future directions include combining LLMs with predictive analytics, integrating them with electronic health records and wearables, and adapting them for multilingual, culturally sensitive dietary guidance. LLMs also hold potential as research and educational tools, assisting in literature synthesis and patient engagement. Their transformative promise depends on cross-disciplinary collaboration, responsible deployment, and clinician training. Ultimately, while LLMs are not a replacement for healthcare professionals, they offer powerful augmentation tools for delivering scalable, personalized, and data-driven nutritional care in an increasingly complex healthcare environment.

Accurate dietary assessment is essential for health research. While smartphone-based food image recognition offers a convenient alternative to traditional methods, existing computer vision approaches struggle with real-world food images and analyze only basic macronutrients, limiting their utility for comprehensive nutritional research. We developed DietAI24, a framework for automated nutrition estimation from food images that combines multimodal large language models (MLLMs) with Retrieval-Augmented Generation (RAG) technology to ground the MLLM’s visual recognition in authoritative nutrition databases rather than relying on the model’s internal knowledge. In our work, we used the Food and Nutrient Database for Dietary Studies (FNDDS) as the authoritative nutrition database. Through this approach, DietAI24 enables accurate nutrient estimation without extensive data collection or model training. DietAI24 significantly outperforms existing methods when evaluated against commercial platforms and computer vision baselines using the ASA24 and Nutrition5k datasets. Performance is measured through mean absolute error (MAE). DietAI24 achieves a 63% reduction in MAE for food weight estimation and four key nutrients and food components compared to existing methods when tested on real-world mixed dishes (p < 0.05). Notably, DietAI24 estimates 65 distinct nutrients and food components, far exceeding the basic macronutrient profiles of existing solutions. DietAI24 demonstrates that integrating MLLMs with RAG and standardized nutrition databases can substantially improve dietary assessment accuracy while enabling comprehensive nutrient analysis. This framework offers a scalable solution for nutrition research and clinical applications, potentially transforming large-scale epidemiological studies and personalized dietary interventions through more accurate and less burdensome dietary data collection. Taking photos of food to track nutrition is convenient, but current apps often guess wrong about what nutrients are in your meals. We created DietAI24, a system that combines artificial intelligence with a trusted nutrition database. When someone takes a food photo, our AI recognizes the foods and looks up their exact nutritional values instead of guessing. We tested DietAI24 against popular nutrition apps using thousands of food images. Our system was 63% more accurate and can identify 65 different nutrients, not just calories and protein, but also important micronutrients like vitamin D, iron, and folate that affect your health. This technology could help researchers better understand diet-related diseases and help doctors give personalized nutrition advice. It makes tracking what you eat easier and more reliable for everyone. Yan et al. present DietAI24, a unified framework that identifies foods, estimates portion sizes, and computes 65 nutrients from real-world food photos using multimodal large language models. This tool outperforms existing methods and standardizes outputs for downstream analysis, enabling efficient, reproducible dietary measurement.

Abstract Transformative change is needed across the food system to improve health and environmental outcomes. As food, nutrition, environmental and health data are generated beyond human scale, there is an opportunity for technological tools to support multifactorial, integrated, scalable approaches to address the complexities of dietary behaviour change. Responsible technology could act as a mechanistic conduit between research, policy, industry and society, enabling timely, informed decision making and action by all stakeholders across the food system. Domain expertise in food, nutrition and health should always be integrated into both the development and continuous deployment of AI‐powered nutritional intelligence (NI) to ensure it is responsible, accurate, safe, useable and effective. Dietary behaviours are complex and improving diet‐related health outcomes requires socio‐cultural‐demographic considerations within the design and deployment of NI tools. This article describes existing examples of NI within the food system and future opportunities. Human‐in‐the‐loop approaches with food, health and nutrition experts involved at all stages including data acquisition, processing, output validation and ongoing quality assurance are essential to ensure evidence‐based practice. The same ethical considerations should apply in this domain as in any other (e.g. privacy, inclusivity, robustness, transparency and accountability) and responsible practice must encompass rigorous standards and alignment with regulatory frameworks. Critical today and in the future is accessibility to appropriate high‐quality food compositional data sets, which include up‐to‐date information on commercially available products that reflect the constantly evolving food landscape to realise the potential of responsible AI to help address the existing food system challenges.

Pretraining domain-specific language models remains an important challenge which limits their applicability in various areas such as agriculture. This paper investigates the effectiveness of leveraging food related text corpora (e.g., food and agricultural literature) in pretraining transformer-based language models. We evaluate our trained language model, called AgriBERT, on the task of semantic matching, i.e., establishing mapping between food descriptions and nutrition data, which is a long-standing challenge in the agricultural domain. In particular, we formulate the task as an answer selection problem, fine-tune the trained language model with the help of an external source of knowledge (e.g., FoodOn ontology), and establish a baseline for this task. The experimental results reveal that our language model substantially outperforms other language models and baselines in the task of matching food description and nutrition.

Abstract Science-informed decisions are best guided by the objective synthesis of the totality of evidence around a particular question and assessing its trustworthiness through systematic processes. However, there are major barriers and challenges that limit science-informed food and nutrition policy, practice, and guidance. First, insufficient evidence, primarily due to acquisition cost of generating high-quality data, and the complexity of the diet-disease relationship. Furthermore, the sheer number of systematic reviews needed across the entire agriculture and food value chain, and the cost and time required to conduct them, can delay the translation of science to policy. Artificial intelligence offers the opportunity to (i) better understand the complex etiology of diet-related chronic diseases, (ii) bring more precision to our understanding of the variation among individuals in the diet-chronic disease relationship, (iii) provide new types of computed data related to the efficacy and effectiveness of nutrition/food interventions in health promotion, and (iv) automate the generation of systematic reviews that support timely decisions. These advances include the acquisition and synthesis of heterogeneous and multimodal datasets. This perspective summarizes a meeting convened at the National Academy of Sciences, Engineering, and Medicine. The purpose of the meeting was to examine the current state and future potential of artificial intelligence in generating new types of computed data as well as automating the generation of systematic reviews to support evidence-based food and nutrition policy, practice, and guidance.

Nutrition plays a pivotal role in preventive health, yet existing digital solutions often lack personalization and accessibility. This study presents an AI-driven framework that integrates machine learning (ML) and natural language processing (NLP) to deliver dynamic, user-centric dietary recommendations. A gradient boosting model, trained on NHANES demographic and anthropometric data, predicts caloric needs with an MAE of 132 kcal, while a locally deployed LLM (Mistral 7B) interprets free-text dietary constraints with 91% accuracy. Rule-based filtering from the USDA database ensures nutritional balance. A pilot usability test (n = 5) confirmed the system’s practicality and satisfaction. The proposed framework addresses key gaps in scalability, privacy, and adaptability, demonstrating the potential of hybrid AI techniques in applied nutrition science. By bridging computational methods with food science, this work offers a reproducible, modular solution for personalized health applications.

BACKGROUND Food categorization and nutrient profiling are labor intensive, time consuming, and costly tasks, given the number of products and labels in large food composition databases and the dynamic food supply. OBJECTIVES This study used a pretrained language model and supervised machine learning to automate food category classification and nutrition quality score prediction based on manually coded and validated data, and compared prediction results with models using bag-of-words and structured nutrition facts as inputs for predictions. METHODS Food product information from University of Toronto Food Label Information and Price Database 2017 (n = 17,448) and University of Toronto Food Label Information and Price Database 2020 (n = 74,445) databases were used. Health Canada's Table of Reference Amounts (TRA) (24 categories and 172 subcategories) was used for food categorization and the Food Standards of Australia and New Zealand (FSANZ) nutrient profiling system was used for nutrition quality score evaluation. TRA categories and FSANZ scores were manually coded and validated by trained nutrition researchers. A modified pretrained sentence-Bidirectional Encoder Representations from Transformers model was used to encode unstructured text from food labels into lower-dimensional vector representations, followed by supervised machine learning algorithms (i.e., elastic net, k-Nearest Neighbors, and XGBoost) for multiclass classification and regression tasks. RESULTS Pretrained language model representations utilized by the XGBoost multiclass classification algorithm reached overall accuracy scores of 0.98 and 0.96 in predicting food TRA major and subcategories, outperforming bag-of-words methods. For FSANZ score prediction, our proposed method reached a similar prediction accuracy (R2: 0.87 and MSE: 14.4) compared with bag-of-words methods (R2: 0.72-0.84; MSE: 30.3-17.6), whereas structured nutrition facts machine learning model performed the best (R2: 0.98; MSE: 2.5). The pretrained language model had a higher generalizable ability on the external test datasets than bag-of-words methods. CONCLUSIONS Our automation achieved high accuracy in classifying food categories and predicting nutrition quality scores using text information found on food labels. This approach is effective and generalizable in a dynamic food environment, where large amounts of food label data can be obtained from websites.

Abstract Objective: To assess the feasibility of using large language models (LLM) to develop research questions about changes to the Special Supplemental Nutrition Program for Women, Infants, and Children (WIC) food packages. Design: We conducted a controlled experiment using ChatGPT-4 and its plugin, MixerBox Scholarly, to generate research questions based on a section of the U.S. Department of Agriculture (USDA) summary of the final public comments on the WIC revision. Five questions weekly for 3 weeks were generated using LLM under two conditions: fed with or without relevant literature. The experiment generated ninety questions, which were evaluated using the Feasibility, Innovation, Novelty, Ethics and Relevance criteria. t tests and multivariate regression examined the difference by feeding status, artificial intelligence model, evaluator and criterion. Setting: The United States. Participants: Six WIC expert evaluators from academia, government, industry and non-profit sectors. Results: Five themes were identified: administrative barriers, nutrition outcomes, participant preferences, economics and other topics. Feeding and non-feeding groups had no significant differences (Coeff. = 0·03, P = 0·52). MixerBox-generated questions received significantly lower scores than ChatGPT (Coeff. = –0·11, P = 0·02). Ethics scores were significantly higher than feasibility scores (Coeff. = 0·65, P < 0·001). Significant differences were found between the evaluators (P < 0·001). Conclusions: The LLM applications can assist in developing research questions with acceptable qualities related to the WIC food package revisions. Future research is needed to compare the development of research questions between LLM and human researchers.

The rising global burden of chronic diseases highlights the limitations of traditional dietary guidelines. Precision Nutrition (PN) aims to deliver personalized dietary advice to optimize individual health, and the effective implementation of PN fundamentally relies on comprehensive and accurate dietary data. However, conventional dietary assessment methods often suffer from quantification errors and poor adaptability to dynamic changes, leading to inaccurate data and ineffective guidance. Machine learning (ML) offers a powerful suite of tools to address these limitations, enabling a paradigm shift across the nutritional management pipeline. Using dietary data as a thematic thread, this article outlines this transformation and synthesizes recent advances across dietary assessment, in-depth mining, and nutritional intervention. Additionally, current challenges and future trends in this domain are also further discussed. ML is driving a critical shift from a subjective, static mode to an objective, dynamic, and personalized paradigm, enabling a loop nutrition management framework. Precise food recognition and nutrient estimation can be implemented automatically with ML techniques like computer vision (CV) and natural language processing (NLP). Integrating with multiple data sources, ML is conducive to uncovering dietary patterns, assessing nutritional status, and deciphering intricate nutritional mechanisms. It also facilitates the development of personalized dietary intervention strategies tailored to individual needs, while enabling adaptive optimization based on users’ feedback and intervention effectiveness. Although challenges regarding data privacy and model interpretability persist, ML undeniably constitutes the vital technical support for advancing PN into practical reality.

The integration of generative artificial intelligence (AI) in poultry nutrition can enhance decision-making, support feed design and improve animal health. This study introduces …

In recent years, major advances in artificial intelligence (AI) have led to the development of powerful AI systems for use in the field of nutrition in order to enhance personalized dietary recommendations and improve overall health and well-being. However, the lack of guidelines from nutritional experts has raised questions on the accuracy and trustworthiness of the nutritional advice provided by such AI systems. This paper aims to address this issue by introducing a novel AI-based nutrition recommendation method that leverages the speed and explainability of a deep generative network and the use of novel sophisticated loss functions to align the network with established nutritional guidelines. The use of a variational autoencoder to robustly model the anthropometric measurements and medical condition of users in a descriptive latent space, as well as the use of an optimizer to adjust meal quantities based on users’ energy requirements enable the proposed method to generate highly accurate, nutritious and personalized weekly meal plans. Coupled with the ability of ChatGPT to provide an unparalleled pool of meals from various cuisines, the proposed method can achieve increased meal variety, accuracy and generalization capabilities. Extensive experiments on 3000 virtual user profiles and 84000 daily meal plans, as well as 1000 real profiles and 7000 daily meal plans, demonstrate the exceptional accuracy of the proposed diet recommendation method in generating weekly meal plans that are appropriate for the users in terms of energy intake and nutritional requirements, as well as the easiness with which it can be integrated into future diet recommendation systems.

Highlights What are the main findings? Large language models generated dietary plans for type 2 diabetes that differed substantially from a guideline-based, dietitian-designed reference diet, particularly in energy intake and dietary fiber adequacy. Most AI-generated diets followed a low-energy, lower-carbohydrate, higher-protein, and insufficient-fiber pattern, with limited evidence of individualized medical nutrition therapy. What are the implications of the main findings? AI-generated dietary plans for diabetes management should not replace professional medical nutrition therapy without expert evaluation. Careful clinical validation and guideline-based refinement are required before integrating large language models into routine diabetes care. Abstract Background/Objectives: Large language models (LLMs) are increasingly used as decision support tools in clinical nutrition, including meal planning for individuals with type 2 diabetes mellitus (T2DM). However, the clinical safety, quantitative accuracy, and guideline adherence of AI-generated dietary plans remain uncertain. This study aimed to evaluate systematic bias and agreement between LLM-generated diets and a guideline-concordant reference diet, and to assess whether current LLMs can function as reliable clinical nutrition decision support tools in T2DM. Methods: Six widely used LLMs generated standardized three-day, 1800 kcal dietary plans for T2DM using an identical prompt. Each day was treated as an independent observation (n = 18). Energy and macronutrient contents were analyzed using professional nutrition software and compared with a dietitian-designed reference diet based on ADA, EASD, IDF, and national guidelines. Agreement was evaluated using Bland–Altman analysis, proportional bias assessment, and intraclass correlation coefficients. Guideline adherence and clinical appropriateness were independently scored by registered dietitians. Results: Most LLM-generated diets systematically deviated from the reference diet, with lower total energy, reduced carbohydrate and fiber content, and variable protein distribution. Bland–Altman analyses demonstrated significant bias and wide limits of agreement for key nutrients, indicating clinically meaningful discrepancies. Guideline adherence scores varied substantially across models, with only one model showing relatively consistent performance. Inter-rater reliability between dietitians was high (ICC = 0.806). Conclusions: Current LLMs exhibit systematic quantitative bias and inconsistent guideline adherence when used for T2DM meal planning. AI-generated dietary plans are not interchangeable with dietitian-guided medical nutrition therapy and may pose clinical risks if used without professional oversight. Careful validation, domain-specific fine-tuning, and integration within supervised clinical workflows are required before implementation in diabetes care.

Background: Precision nutrition (PN) aims to personalize dietary guidance by accounting for inter-individual variability across biological, metabolic, lifestyle, and environmental factors influencing nutritional needs and health outcomes. While traditional Artificial Intelligence (AI) has advanced nutritional research through systems like automated dietary assessment, these models often operate rigidly. Generative AI (GenAI) introduces the capacity for adaptive interventions for enhanced PN. However, the scope and maturity of its applications remain insufficiently characterized. Objective: This review examined original works applying GenAI in PN, focusing on application, methodology, and limitations. Methods: A systematic search was conducted in PubMed, ACM Digital Library, and Scopus. Inclusion criteria focused on original works deploying GenAI models in PN contexts. Included works were further formally assessed based on data used, validation, transparency, bias, and security and privacy. Results: 21 eligible studies were identified, all published after 2024. The literature indicated a surge in large language model-based systems for personalized dietary recommendations, followed by applications in data foundation building and food effect understanding. A recurrent limitation was questionable evaluation on synthetic data and hallucinations, necessitating a human-expert-in-the-loop, especially in high-stakes clinical settings. Additionally, only 4 of 21 reviewed studies incorporated biological content or biological inputs, and fewer approached biologically grounded PN within implemented personalization workflows using metabolic and/or genomic variables. Conclusions: Although GenAI research in PN is expanding rapidly, most applications remain personalized at a user-preference level rather than including biological determinants. The need for standardized reporting, stronger genome-informed modeling, and consistent human-in-the-loop validation protocols is further highlighted to advance towards holistic PN.

… 摘 要: 营养信息学正由传统基于规则与常规机器学习范式,迈向以大语言模型(large language model,LLM)与多 模态大模型(multimodal large language models,MLLM)为核心的新阶段.本文…

Abstract Background The integration of Large Language and Vision Assistant models with food and nutrition data enables multimodal meal analysis and contextual dietary guidance. Despite this potential, the reliability and practical usefulness of such systems for supporting everyday dietary decision-making remain underexplored. Objective This study introduces Purrfessor, an innovative artificial intelligence (AI) chatbot designed to provide personalized dietary guidance through interactive, multimodal engagement. The study aimed to evaluate its performance in ingredient recognition and recipe generation. Methods The Purrfessor chatbot was trained using a combination of the FoodData Central database from the US Department of Agriculture (USDA), the Recipe2img dataset featuring food images and corresponding recipes, a curated human-annotated dataset derived from Recipe1M, and a customized question-and-answer dialogue dataset. The system operates under a session-based, multiturn interaction paradigm, with memory retained only within an active session and no cross-session memory persistence. We implemented a 2-phase evaluation framework combining AI-based performance assessment and human scoring. Results Purrfessor achieved a high average cosine similarity of 0.90 in ingredient recognition with human-coded references. In GPT-4.1–based (OpenAI) evaluation of recipe generation quality, Purrfessor outperformed the raw Large Language and Vision Assistant model across all evaluated dimensions, with the largest improvements in completeness (7.44 vs 6.52), consistency (8.90 vs 7.81), and clarity (9.13 vs 8.39). Overall recipe quality improved from 7.66 to 8.35. Automatic metrics indicated strong ingredient coverage (0.78) and moderate step complexity (0.74), with lower coherence (0.62) and temperature and time specification (0.59), yielding an overall structured score of 0.68. Human evaluators rated Purrfessor’s question-and-answer accuracy highly: correctness (mean 8.71, SD 1.15), relevance (mean 9.99, SD 0.10), and clarity (mean 9.33, SD 0.68). Error analysis indicated that 56% of responses contained minor hallucinations (ie, inclusion of inferred secondary details or invisible garnishes). At the same time, core food identification and overall recipe logic remained accurate. Conclusions Findings highlight the role of anthropomorphic chatbot design and multimodal AI in supporting engaging dietary health conversations. This study offers an example of AI-driven, evidence-based dietary guidance and underscores the potential of health chatbots to nudge informed health decision-making. Insights contribute to the development of digital health interventions and personalized health communication strategies, with implications for the design of engaging, user-centered AI health assistants.

Accurately monitoring diet remains a challenge, especially for home-cooked or restaurant meals that do not carry nutritional labels and are difficult to record using conventional calorie-counting apps. Most existing tools rely on manual entry or barcode scanning, which are time-consuming and often inaccurate. To address these limitations, this study introduces an AI-powered calorie tracking system that combines image captioning (BLIP model) with large language models (Groq LLM via LangChain) to automatically recognize food items, estimate their nutritional values, and generate personalized feedback. Users can simply upload a photo of a meal or enter food names, and the system provides calorie and nutrient breakdowns, daily summaries, and fitness suggestions when caloric excess is detected. Reports are automatically compiled into a downloadable PDF for long-term monitoring. The system demonstrated strong performance, achieving a precision of 92.4%, a recall of 91.1%, and an F1-score of 91.7% in food recognition, with an AUC of 0.95 for nutritional estimation. Case studies confirmed its ability to handle multi-item meals, correctly identify key nutrients, and provide contextual health advice. To ensure transparency and trust, the framework incorporates Explainable AI (XAI) techniques such as Grad-CAM, SHAP, LIME, and EDGE, which help users and healthcare professionals understand how predictions are generated. By reducing manual effort, adapting to food variability, and offering real-time, interpretable insights, this system provides a practical, reliable, and user-friendly solution for personal health management, chronic disease care, and commercial fitness applications.

Abstract Background Traditional Chinese Medicine (TCM) emphasizes the concept of medicine food homology (MFH), which integrates dietary therapy into health care. However, the practical application of MFH principles relies heavily on expert knowledge and manual interpretation, posing challenges for automating MFH-based dietary recommendations. Although large language models (LLMs) have shown potential in health care decision support, their performance in specialized domains such as TCM is often hindered by hallucinations and a lack of domain knowledge. The integration of uncertain knowledge graphs (UKGs) with LLMs via retrieval-augmented generation (RAG) offers a promising solution to overcome these limitations by enabling a structured and faithful representation of MFH principles while enhancing LLMs’ ability to understand the inherent uncertainty and heterogeneity of TCM knowledge. Consequently, it holds potential to improve the reliability and accuracy of MFH-based dietary recommendations generated by LLMs. Objective This study aimed to introduce Yaoshi-RAG, a framework that leverages UKGs to enhance LLMs' capabilities in generating accurate and personalized MFH-based dietary recommendations. Methods The proposed framework began by constructing a comprehensive MFH knowledge graph (KG) through LLM-driven open information extraction, which extracted structured knowledge from multiple sources. To address the incompleteness and uncertainty within the MFH KG, UKG reasoning was used to measure the confidence of existing triples and to complete missing triples. When processing user queries, query entities were identified and linked to the MFH KG, enabling retrieval of relevant reasoning paths. These reasoning paths were then ranked based on triple confidence scores and entity importance. Finally, the most informative reasoning paths were encoded into prompts using prompt engineering, enabling the LLM to generate personalized dietary recommendations that aligned with both individual health needs and MFH principles. The effectiveness of Yaoshi-RAG was evaluated through both automated metrics and human evaluation. Results The constructed MFH KG comprised 24,984 entities, 22 relations, and 29,292 triples. Extensive experiments demonstrate the superiority of Yaoshi-RAG in different evaluation metrics. Integrating the MFH KG significantly improved the performance of LLMs, yielding an average increase of 14.5% in Hits@1 and 8.7% in F1-score, respectively. Among the evaluated LLMs, DeepSeek-R1 achieved the best performance, with 84.2% in Hits@1 and 71.5% in F1-score, respectively. Human evaluation further validated these results, confirming that Yaoshi-RAG consistently outperformed baseline models across all assessed quality dimensions. Conclusions This study shows Yaoshi-RAG, a new framework that enhances LLMs’ capabilities in generating MFH-based dietary recommendations through the knowledge retrieved from a UKG. By constructing a comprehensive TCM knowledge representation, our framework effectively extracts and uses MFH principles. Experimental results demonstrate the effectiveness of our framework in synthesizing traditional wisdom with advanced language models, facilitating personalized dietary recommendations that address individual health conditions while providing evidence-based explanations.

… Based on the test results, the best-performing LLM was selected as the base model. … accuracy of dietary recommendations generated by the LLM. This involved dense retrieval from a …

The concept of medicine and food homology in traditional Chinese medicine (TCM) emphasized the dual role of certain material as both food and medicine, offering nutritional and therapeutic benefits. Edible herbal formulas, derived from this principle, are valuable for health management and chronic disease prevention. This study proposes a domain-specific prescription recommendation model enriched by TCM edible herbal formula knowledge called TCM-DS model. A dataset including symptoms, TCM constitutions, formulas and their corresponding ingredients was developed. DeepSeek R1 base model was fine-tuned utilizing Low-rank adaptation (LoRA) fine-tuning and a retrieval-augmented generation (RAG) module to increase recommendation accuracy. TCM-DS model was evaluated against general-purpose large language models. The proposed TCM-DS model demonstrated superior performance, achieving a recommendation precision of 0.9924. Comparative experiments showed its robustness, with the highest precision scores for both forward and reverse symptom sequences compared with general-purpose large language models. A user-friendly platform was developed based on TCM-DS model, enabling automated constitution analysis and personalized formula recommendations. In conclusion, we proposed an intelligent TCM edible herbal formula recommendation model called TCM-DS. Its accompanying platform automated constitution identification and formula recommendation, advancing intelligent applications in TCM practice.

… To identify the most suitable LLM for generating dietary recommendations based on the TCM principle of “One Root of Medicine and Food”, we designed a unified set of prompts to …

Background Potato is the fourth largest food crop in the world, but potato cultivation faces serious threats from various diseases and pests. Despite significant advancements in research on potato disease resistance, these findings are scattered across numerous publications. For researchers, obtaining relevant knowledge by reading and organizing a large body of literature is a time-consuming and labor-intensive process. Therefore, systematically extracting and organizing the relationships between potato genes and diseases from the literature to establish a potato gene-disease knowledge base is particularly important. Unfortunately, there is currently no such gene-disease knowledge base available. Methods In this study, we constructed a Potato Gene-Disease Knowledge Base (PotatoG-DKB) using natural language processing techniques and large language models. We used PubMed as the data source and obtained 2,906 article abstracts related to potato biology, extracted entities and relationships between potato genes and related disease, and stored them in a Neo4j database. Using web technology, we also constructed the Potato Gene-Disease Knowledge Portal (PotatoG-DKP), an interactive visualization platform. Results PotatoG-DKB encompasses 22 entity types (such as genes, diseases, species, etc.) of 5,206 nodes and 9,443 edges between entities (for example, gene-disease, pathogen-disease, etc.). PotatoG-DKP can intuitively display associative relationships extracted from literature and is a powerful assistant for potato biologists and breeders to understand potato pathogenesis and disease resistance. More details about PotatoG-DKP can be obtained at https://www.potatogd.com.cn/.

OBJECTIVE Despite adequate dialysis, the prevalence of hyperkalemia in Chinese hemodialysis(HD) patients remains elevated. This study aims to evaluate the effectiveness of a dietary recommendation system driven by Generative Pre-trained Transformers (GPTs) in managing potassium levels in HD patients. METHODS We implemented a bespoke dietary guidance tool utilizing GPTs technology. Patients undergoing HD at our center were enrolled for the study from October 2023 to November 2023. The intervention comprised two distinct phases. Initially, patients were provided with conventional dietary education focused on potassium management in HD. Subsequently, in the second phase, they were introduced to a novel GPT-based dietary guidance tool. This AI-powered tool offered real-time insights into the potassium content of various foods and personalized dietary suggestions. The effectiveness of the AI tool was evaluated by assessing the precision of its dietary recommendations. Additionally, we compared pre-dialysis serum potassium levels and the proportion of patients with hyperkalemia among patients before and after the implementation of the GPT-based dietary guidance system. RESULTS In our analysis of 324 food photographs uploaded by 88 HD patients, the GPTs system evaluated potassium content with an overall accuracy of 65%. Notably, the accuracy was higher for high-potassium foods at 85%, while it stood at 48% for low-potassium foods. Furthermore, the study examined the effect of GPTs-based dietary advice on patients' serum potassium levels, revealing a significant reduction in those adhering to GPTs recommendations compared to recipients of traditional dietary guidance (4.57±0.76 mmol/L vs. 4.84±0.94 mmol/L, p = 0.004). Importantly, Compared to traditional dietary education, dietary education based on the GPTs tool reduced the proportion of hyperkalemia in HD patients from 39.8% to 25%(p=0.036). CONCLUSION These results underscore the promising role of AI in improving dietary management for HD patients. Nonetheless, the study also points out the need for enhanced accuracy in identifying low potassium foods. It paves the way for future research, suggesting the incorporation of extensive nutritional databases and the assessment of long-term outcomes. This could potentially lead to more refined and effective dietary management strategies in HD care.

ABSTRACT In the era of Industry 4.0, where automation and digitalization are central to processes and systems, artificial intelligence (AI) is becoming an essential tool that provides innovative solutions across various fields. Nutrition, as a vital component of public health, is among the areas increasingly shaped by the integration of AI technologies. This review aims to identify the strengths and limitations of these models, assess their potential as decision-support tools for healthcare professionals, and shed light on the existing gaps in the field. A comprehensive literature search was conducted using PubMed, Scopus, Web of Science, and Google Scholar databases to identify relevant studies published between January 2019 and May 2025. A total of 44 articles were included in the review. The main findings suggest that the application of AI in nutrition is still emerging, with the majority of studies centered on dietary assessment. There is comparatively less emphasis on food estimation, disease prediction, lifestyle interventions, and understanding diet-related diseases. Further clinical studies are necessary to evaluate the effectiveness of AI-based interventions.

People increasingly prioritize a balanced diet to enhance well-being, yet making informed dietary choices remains challenging amidst the abundance of options. To address this, we developed a meal image recognition and healthy meal combination recommender system, integrating generative artificial intelligence (AI). Convolutional neural networks were used for precise meal image recognition to identify diverse food items. The generative AI-augmented recommendation engine offers personalized meal suggestions aligned with nutritional goals and dietary preferences, utilizing a nutrition knowledge base to ensure overall well-being. The system's feasibility was validated, illustrating its excellence in meal recognition accuracy, recommendation diversity, and user engagement. By integrating generative AI, the system shows its potential to enhance dietary recommendations and public health.

… GenAI in delivering dietary recommendations across multiple … The advent of artificial intelligence (AI)-based chatbots has … to deliver personalized dietary recommendations. The …

Personalized dietary guidance has become an essential requirement in modern healthcare due to increasing lifestyle-related health issues. This paper presents NutriSmart, an intelligent web-based nutrition advisory system that delivers personalized dietary recommendations by integrating machine learning and generative artificial intelligence. The proposed solution employs a three-tier architecture consisting of a dynamic, conversational user interface and a backend processing layer powered by machine learning model and a large language model (LLM). Nutritional data are sourced from the USDA FoodData Central Foundation Foods database, which provides comprehensive information on food composition and nutrient profiles. A nutrient KNN-cluster scoring mechanism is used to evaluate and assign food recommendations based on individual user requirements. The machine learning model generates initial dietary predictions, which are subsequently combined with contextual prompts and passed to Google Generative AI to produce real-time, personalized meal recommendations. Designed with a conversational interaction paradigm, NutriSmart enables users to query dietary needs naturally and receive tailored advice in an intuitive and user-friendly manner. By combining data-driven prediction with generative reasoning, the proposed system demonstrates potential as a scalable and intelligent nutrition guidance platform for personalized health and diet management.

AbstractThis study presents a comparative analysis of rule-based and generative AI-based diet and workout recommendation systems, highlighting their effectiveness in delivering personalized health guidance. The rule-based model relies on predefined logic and static conditions, whereas the generative system leverages OpenAI’s language models via Lang Chain to generate dynamic, context-aware recommendations based on user inputs. To augment this comparison with cultural richness and holistic understanding, aspects of Indian Knowledge Systems (IKS) specifically Ayurveda and indigenous fitness practices such as Yoga are brought in as a complementary model. Prakriti (constitution of the body), Ritucharya (seasonal adjustment), and food-mind typologies (Satvik, Rajasik, Tamasik) form a basis for culturally sensitive customization. The blending of IKS has the potential to improve both systems by infusing conventional health knowledge into contemporary AI-based recommendation systems, thus enhancing relevance, acceptability, and user satisfaction in multicultures.

We propose a recommender system framework that uses Generative AI and family health profiling to transform grocery shopping into a precision health intervention. The framework enables real-time nudges, personalized scoring, and dynamically generated ingredient transparency using large language models (LLMs). We implement a practical ingredient classifier with a ChatGPT-family model accessed via the OpenAI API and expose its decisions through a web front end. This vision lays the groundwork for proactive nutrition coaching embedded within everyday consumer behavior, offering scalable, ethical, and equity-aware deployment across retail ecosystems. Beyond technical feasibility, we discuss implications for healthcare, retail, and society, and outline a prototype with illustrative telemetry. To set scope, we present a prototype with mock telemetry and illustrative evaluation rather than large-scale deployment, and we discuss limitations such as partial JSON-constrained outputs and reliance on external nutrition databases.

… To quantify the amount of a given nutrient in a food item and determine its suitability based on dietary recommendations, we use fuzzy logic to categorise nutrient quantity as low, …

This study addresses the limitations of traditional prescribed meal plans, which lack personalization and often prove monotonous and challenging for patients to follow. We propose a novel approach employing generative artificial intelligence in the context of a learning health system, with an emphasis on inpatient clinical dietetics. The system incorporates two key models: the Meal Plan Generation Model, MeaIGM, and the Meal Plan Image Generation Model, MealImageGM, leveraging state-of-the-art large language models. Patient information from electronic health records and clinical dietetics guidelines are incorporated into prompts for MeaIGM, which is refined through nutritionist validations and users' feedback. On the other hand, MealImageGM generates visual representations of meal plans to enhance patient engagement, utilizing crowd-sourced feedback to optimize image generation prompts. The overall system process includes extracting data from electronic health records, pre-designed user meal generation prompts, and the generation of personalized meal plans and images. Nutritionists play a crucial role in monitoring patient adherence and preferences, contributing to a continuous learning health system cycle. The proposed framework ensures clinically appropriate and personalized meal plans, aligning with dynamic dietary recommendations. The study emphasizes the importance of patient-physician co-creation for constant optimization and highlights the potential positive impact on health outcomes.

… generative AI-based health advisory system specifically designed to deliver real-time, personalized recommendations … being used to create personalized diet plans. They estimate daily …

… While generative AI excels in generating coherent and … to inaccuracies or unsafe dietary recommendations. Their outputs … Our work extends the use of generative AI to output recipes …

Background/Objectives: Artificial intelligence (AI) has emerged as a transformative force in healthcare, with nutrition and dietetics becoming key areas of application. AI technologies are being employed to enhance dietary assessment, personalize nutrition plans, manage chronic diseases, deliver virtual coaching, and support public health nutrition. This review aims to critically synthesize the current literature on AI applications in nutrition, identify research gaps, and outline directions for future development. Methods: A systematic literature search was conducted across PubMed, Scopus, Web of Science, and Google Scholar for peer-reviewed publications from January 2020 to July 2025. The search included studies involving AI applications in nutrition, dietetics, or public health nutrition. Articles were screened based on predefined inclusion and exclusion criteria. Thematic analysis grouped findings into six categories: dietary assessment, personalized nutrition and chronic disease management, generative AI and conversational agents, global/public health nutrition, sensory science and food innovation, and ethical and professional considerations. Results: AI-driven systems show strong potential for improving dietary tracking accuracy, generating personalized diet recommendations, and supporting disease-specific nutrition management. Chatbots and large language models (LLMs) are increasingly used for education and support. Despite this progress, challenges remain regarding model transparency, ethical use of health data, limited generalizability across diverse populations, and underrepresentation of low-resource settings. Conclusions: AI offers promising solutions to modern nutritional challenges. However, responsible development, ethical oversight, and inclusive validation across populations are essential to ensure equitable and safe integration into clinical and public health practice.

Accurate estimation of nutritional information from food images remains a challenging problem. Most existing approaches rely on deep image models fine-tuned with extensive food annotations or require detailed user inputs (e.g., portion size, cooking method), both of which are prone to error. Motivated by the workflow of nutrition experts, we propose a two-step prompting framework leveraging off-the-shelf Multimodal Large Language Models (MLLMs). The first step deconstructs the dish into its components listing major ingredients, portion sizes, and cooking details while the second step computes total calories and macronutrients. This approach alleviates the need for heavy fine-tuning or large ingredient databases, by instead harnessing the compositional reasoning capabilities of general MLLMs. We evaluate the method on both a subset of the Nutrition5k dataset (Nutrition320) and real-world samples from the Gindee application (Gindee121), achieving more accurate estimates than one-step direct queries. Additional experiments with visual prompts (bounding boxes, segmentation masks) further demonstrate the robustness and adaptability of our approach. Notably, our findings reveal that guiding MLLMs through a structured two-step reasoning process–separating “what is on the plate” from “how it translates nutritionally”–substantially improves the reliability of image-based macronutrient estimation.

The evolution of food products has become rapid with time. There is a significant rise in the consumption of packaged food items, which has slowly started to become a concern for consumer health. Understanding the ingredients present in the packaged food is of utmost importance-but it is often overlooked. Consumption of such food items may lead to severe health complications to the consumers posed by uninformed consumption (e.g., allergens, additives, dietary mismatches). This paper proposes a solution to this problem-food ingredient scanner, that analyses the ingredients mentioned on the label of packaged food items and provides output to the consumers accordingly, helping them to understand the complex term easily. Through the combination of Text Extraction, Optical Character Recognition and Large Language Model, the system acts as a digital nutrition advising assistant that provides an insight of the ingredients-which may be necessary for the consumers to know. The system also includes a risk assessment module that highlights the potential health risks, allergens and dietary restrictions/considerations. Currently, a Minimum Viable Product (MVP) is developed to implement the core features. This result highlights the potential of the tool, to bridge the gap between the consumer’s lack of understanding of food ingredients that lead them to confusion, with a system that helps them understand the ingredients and provide valuable information that will provide them the insights, helping them decide whether to consume the packaged food item or not. This solution not only solves the issue, but also helps promote greater awareness of food safety to the society. This work also demonstrates how AI can enhance public health, by making food labels more accessible and comprehensive.

Chronic diseases linked to poor dietary choices necessitate innovative tools for personalized nutrition management. This paper presents NutriGuard, a multimodal AI framework that integrates optical character recognition (OCR), deep learning, and fine-tuned large language models (LLMs) to deliver realtime, context-aware dietary recommendations tailored to users' health profiles. The system begins by extracting nutritional information from food labels via a hybrid OCR pipeline, combining PaddleOCR for English and Surya for Bengali text. For lowquality images, a convolutional neural network (CNN) trained on a dataset of food labels detects product categories, crossreferencing a nutritional database to fill information gaps. A Llama-3.2 model, fine-tuned on clinical guidelines and medical literature, analyzes extracted data against user-specific health conditions (e.g., diabetes, hypertension) to generate risk assessments, substitution suggestions, and personalized meal plans. This work advances AI-driven preventive healthcare by establishing a multilingual, clinically validated framework for dietary risk mitigation. The system's modular design permits rapid adaptation to regional food cultures and emerging nutritional research, addressing critical gaps in scalable personalized nutrition management. Future integration with continuous glucose monitors and gut microbiome data promises to enable dynamic, biomarker-informed dietary optimization.

Scientific evidence is fundamental to uncovering insights about health, including food and nutritional claims. Substantiating such claims requires robust scientific procedures that often include clinical studies, biochemical analyses, and the examination of multiple forms of data. The growing capabilities of artificial intelligence (AI) and large language models (LLMs) present new opportunities for analyzing food–health relationships and supporting health claim validation. Yet, applying these technologies to the food and nutrition domain raises challenges that differ from those encountered in broader biomedical text mining (TM). In this perspective, we review key issues, including the complexity and heterogeneity of food-related data, the scarcity of food-specific language models and standardized resources, difficulties in interpreting nuanced and often contradictory evidence, and requirements for integrating AI tools into regulatory workflows. We compare modern LLM approaches with traditional TM methods and discuss how each may complement the other. Our position is that, despite their promise, current AI and LLM tools cannot yet reliably handle the subtleties of food–health evidence without substantial domain-specific refinement and human expert oversight. We advocate for hybrid approaches that combine the precision of established TM techniques with the analytical breadth of LLMs, supported by harmonized ontologies, multidimensional evaluation frameworks, and human-in-the-loop validation, particularly in regulatory contexts. We also highlight the importance of public education, transparent communication standards, and coordinated cross-disciplinary efforts to ensure these technologies serve broader goals of food safety, consumer trust, and global health.

Background/Objectives: The rapidly expanding landscape of digital technologies is transforming innovation processes across industries, and the food sector is increasingly encouraged to adopt novel tools that can enhance development workflows and support competitive positioning. In the context of Industry 4.0, it is particularly important to examine open innovation approaches that may increase the efficiency of engineers and researchers involved in the research and development of food supplements. Such approaches enable broader access to relevant scientific information, including new bioactive ingredient research and their physiological implications, potentially contributing to the development of better-informed and higher-quality products. Methods: In the present study, we evaluated the deep research capabilities of several popular large language models to assess their suitability for supporting the conceptual design of a blood glucose-optimizing food supplement intended for prediabetes management. The comparative analysis focused on the level of detail in the outputs generated by each model, the robustness of the conclusions drawn, and the capacity to produce formulation-oriented recommendations grounded in scientific literature and regulatory frameworks. Our evaluation was primarily qualitative and subjective, highlighting both the potential and limitations of these models. Moreover, the study outlines a forward-looking concept for product validation using wearable smart devices and medically certified wearable devices with continuous biometric monitoring, which could provide an innovative avenue for assessing supplement efficacy. Results: The findings indicate that large language models can support the collection, organization, and preliminary interpretation of complex scientific information. Conclusions: Nevertheless, expert input remains essential for accurate evaluation, scientific validation, and regulatory compliance, as these models cannot yet replace domain expertise or rigorous experimentation in food supplement development.

Objective This study evaluated the validity and reliability of large language model (LLM) responses on dietary supplements (DS), a domain marked by scientific controversy and misinformation. The goal was to support informed consumer decisions and guide improvements in LLM performance. Methods We collected responses from GPT-4 and GPT-4o on the effects of 30 DS on six diseases. Two medical professionals categorized each response as “Effective,” “Uncertain,” or “Not Effective.” They also created a guideline to assess evidence-based effectiveness and compared it with LLM-generated responses to determine accuracy. Additionally, we conducted qualitative content analysis to identify response patterns and misleading content. Results GPT-4 and GPT-4o affirmed DS effectiveness in only 10% of cases, with 40% rated as “Uncertain” and 50% as “Not Effective.” Accuracy was about 57%, considerably lower than that observed in nutrition-related studies (57% in DS vs. 80% ∼ in structured nutrition tasks”). Content analysis showed templated responses, frequent ambiguity, and occasional inclusion of irrelevant or incorrect information. Conclusion Our findings suggest that ChatGPT's responses on dietary supplements are generally cautious but often ambiguous, with a moderate risk of misinformation. As generative AI becomes a common source for health advice, these limitations could mislead users. Enhancing LLMs' evidence-based accuracy and ensuring consistent professional guidance are essential. Innovation This is the first study to assess the validity and reliability of LLM-generated responses on dietary supplements using both quantitative and qualitative methods. We also developed a novel evidence-based framework to evaluate supplement effectiveness, providing a new tool for future research and supporting safer AI-assisted health communication.

… In the United States, there is no legal definition of “functional food”; instead, foods are … In this study, we used the LLM Gemini API to compare and analyze major health food regulatory …

A centralized system that along with monitoring your food consumption, also provides you with intelligent consultations and advice is a crucial need in the times of digital medicine. This paper reflects the idea and execution of a complete end-to-end and seamless food diary and chatbot system where users can track their multiple nutrient consumption by inculcating to it their meals, and interact with a smart chatbot which can use these information and with the help of AI, provide appropriate advice. This application is developed with a modern web development stack and is very dynamic, begineer-friendly and scalable for self nutritional tracking. Including Biomistral's large language model (LLM) also optimizes user experience and engage context-full conversations. The paper addresses the system design, implementation, evaluation and servey of relevant works and technologies in this field.

… LLM-based conversational agent called E-Mealio. Its objective is to help users better understand the impact of food … and used to improve future recommendations and dietary analysis. …

This paper focuses on implementing and evaluating a Retrieval-Augmented Generation (RAG) system that provides personalized fitness and dietary guidance by combining real-time data from wearable devices with curated knowledge sources. While the current implementation depends on a carefully selected set of PDF files as a source of knowledge base, extracting relevant information in order to frame the answer to a question by the user, the planned integration of a curated, well-structured database aims to enhance the system’s reliability and scope. The paper presents an overview of RAG-based system-related literature, describes the process of implementing the RAG component, and discusses the effectiveness in generating relevant and accurate answers. This implementation demonstrates the capability of RAG technology to develop interactive and effective solutions for dietary and fitness applications and sets a foundation for further system improvements.

… Dietary supplements (DS) are products intended to add nutritional value to the diet, often … .0 knowledge base, the RAG framework informs the user that the specific knowledge is not …

Background Cardiovascular disease (CVD) remains the leading cause of death worldwide, yet many web-based sources on cardiovascular (CV) health are inaccessible. Large language models (LLMs) are increasingly used for health-related inquiries and offer an opportunity to produce accessible and scalable CV health information. However, because these models are trained on heterogeneous data, including unverified user-generated content, the quality and reliability of food and nutrition information on CVD prevention remain uncertain. Recent studies have examined LLM use in various health care applications, but their effectiveness for providing nutrition information remains understudied. Although retrieval-augmented generation (RAG) frameworks have been shown to enhance LLM consistency and accuracy, their use in delivering nutrition information for CVD prevention requires further evaluation. Objective To evaluate the effectiveness of off-the-shelf and RAG-enhanced LLMs in delivering guideline-adherent nutrition information for CVD prevention, we assessed 3 off-the-shelf models (ChatGPT-4o, Perplexity, and Llama 3-70B) and a Llama 3-70B+RAG model. Methods We curated 30 nutrition questions that comprehensively addressed CVD prevention. These were approved by a registered dietitian providing preventive cardiology services at an academic medical center and were posed 3 times to each model. We developed a 15,074-word knowledge bank incorporating the American Heart Association’s 2021 dietary guidelines and related website content to enhance Meta’s Llama 3-70B model using RAG. The model received this and a few-shot prompt as context, included citations in a Context Source section, and used vector similarity to align responses with guideline content, with the temperature parameter set to 0.5 to enhance consistency. Model responses were evaluated by 3 expert reviewers against benchmark CV guidelines for appropriateness, reliability, readability, harm, and guideline adherence. Mean scores were compared using ANOVA, with statistical significance set at P<.05. Interrater agreement was measured using the Cohen κ coefficient, and readability was estimated using the Flesch-Kincaid readability score. Results The Llama 3+RAG model scored higher than the Perplexity, GPT-4o, and Llama 3 models on reliability, appropriateness, guideline adherence, and readability and showed no harm. The Cohen κ coefficient (κ>70%; P<.001) indicated high reviewer agreement. Conclusions The Llama 3+RAG model outperformed the off-the-shelf models across all measures with no evidence of harm, although the responses were less readable due to technical language. The off-the-shelf models scored lower on all measures and produced some harmful responses. These findings highlight the limitations of off-the-shelf models and demonstrate that RAG system integration can enhance LLM performance in delivering evidence-based dietary information.

… RAG model in this study were rigorously curated from diverse sources, including federal agencies and scientific publications, to establish a comprehensive knowledge base for nutrition-…

Artificial intelligence has driven significant progress in the nutrition field, especially through multimedia-based automatic dietary assessment. However, existing automatic dietary assessment systems often overlook critical non-visual factors, such as ingredient substitutions that can significantly alter nutritional content, and rarely account for individual dietary needs, including allergies, restrictions, and preferences. In Switzerland, while food-related information is available, it remains fragmented, and no centralized repository integrates all relevant nutrition-related aspects within a Swiss context. To bridge this gap, we introduce the Swiss Food Knowledge Graph (SwissFKG), the first resource, to our best knowledge, to ever unite recipes, ingredients, and their substitutions with nutrient data, dietary restrictions, allergen information, and national nutrition guidelines under one graph. We establish a Large Language Model (LLM)-powered enrichment pipeline for populating the graph, whereby we further present the first benchmark of four off-the-shelf (<70 B parameter) LLMs for food knowledge augmentation. Our results demonstrate that LLMs can effectively enrich the graph with relevant nutritional information. Our SwissFKG goes beyond recipe recommendations by offering ingredient-level information such as allergen and dietary restriction information, and guidance aligned with nutritional guidelines. Moreover, we implement a Graph-Retrieval Augmented Generation (Graph-RAG) application to showcase how the SwissFKG's rich natural-language data structure can help LLM answer user-specific nutrition queries, and we evaluate LLM-embedding pairings by comparing user-query responses against predefined expected answers. As such, our work lays the foundation for the next generation of dietary assessment tools that blend visual, contextual, and cultural dimensions of eating.

Recipes available on the web often lack nutritional transparency and clear indicators of dietary suitability. While searching by title is straightforward, exploring recipes that meet combined dietary needs, nutritional goals, and ingredient-level preferences remains challenging. Most existing recipe search systems do not effectively support flexible multi-dietary reasoning in combination with user preferences and restrictions. For example, users may seek gluten-free and dairy-free dinners with suitable substitutions, or compound goals such as vegan and low-fat desserts. Recent systematic reviews report that most food recommender systems are content-based and often non-personalized, with limited support for dietary restrictions, ingredient-level exclusions, and multi-criteria nutrition goals. This paper introduces DietQA, an end-to-end, language-adaptable chatbot system that integrates a Knowledge Graph (KG), Retrieval-Augmented Generation (RAG), and a Large Language Model (LLM) to support personalized, dietary-aware recipe search and question answering. DietQA crawls Greek-language recipe websites to extract structured information such as titles, ingredients, and quantities. Nutritional values are calculated using validated food composition databases, and dietary tags are inferred automatically based on ingredient composition. All information is stored in a Neo4j-based knowledge graph, enabling flexible querying via Cypher. Users interact with the system through a natural language chatbot friendly interface, where they can express preferences for ingredients, nutrients, dishes, and diets, and filter recipes based on multiple factors such as ingredient availability, exclusions, and nutritional goals. DietQA supports multi-diet recipe search by retrieving both compliant recipes and those adaptable via ingredient substitutions, explaining how each result aligns with user preferences and constraints. An LLM extracts intents and entities from user queries to support rule-based Cypher retrieval, while the RAG pipeline generates contextualized responses using the user query and preferences, retrieved recipes, statistical summaries, and substitution logic. The system integrates real-time updates of recipe and nutritional data, supporting up-to-date, relevant, and personalized recommendations. It is designed for language-adaptable deployment and has been developed and evaluated using Greek-language content. DietQA provides a scalable framework for transparent and adaptive dietary recommendation systems powered by conversational AI.

Background Dietary supplements (DSs) are widely used to improve health and nutrition, but challenges related to misinformation, safety, and efficacy persist due to less stringent regulations compared with pharmaceuticals. Accurate and reliable DS information is critical for both consumers and health care providers to make informed decisions. Objective This study aimed to enhance DS-related question answering by integrating an advanced retrieval-augmented generation (RAG) system with the integrated Dietary Supplement Knowledgebase 2.0 (iDISK2.0), a dietary supplement knowledge base, to improve accuracy and reliability. Methods We developed iDISK2.0 by integrating updated data from authoritative sources, including the Natural Medicines Comprehensive Database, the Memorial Sloan Kettering Cancer Center database, Dietary Supplement Label Database, and Licensed Natural Health Products Database, and applied advanced data cleaning and standardization techniques to reduce noise. The RAG system combined the retrieval power of a biomedical knowledge graph with the generative capabilities of large language models (LLMs) to address limitations of stand-alone LLMs, such as hallucination. The system retrieves contextually relevant subgraphs from iDISK2.0 based on user queries, enabling accurate and evidence-based responses through a user-friendly interface. We evaluated the system using true-or-false and multiple-choice questions derived from the Memorial Sloan Kettering Cancer Center database and compared its performance with stand-alone LLMs. Results iDISK2.0 integrates 174,317 entities across 7 categories, including 8091 dietary supplement ingredients; 163,806 dietary supplement products; 786 diseases; and 625 drugs, along with 6 types of relationships. The RAG system achieved an accuracy of 99% (990/1000) for true-or-false questions on DS effectiveness and 95% (948/100) for multiple-choice questions on DS-drug interactions, substantially outperforming stand-alone LLMs like GPT-4o (OpenAI), which scored 62% (618/1000) and 52% (517/1000) on these respective tasks. The user interface enabled efficient interaction, supporting free-form text input and providing accurate responses. Integration strategies minimized data noise, ensuring access to up-to-date, DS-related information. Conclusions By integrating a robust knowledge graph with RAG and LLM technologies, iDISK2.0 addresses the critical limitations of stand-alone LLMs in DS information retrieval. This study highlights the importance of combining structured data with advanced artificial intelligence methods to improve accuracy and reduce misinformation in health care applications. Future work includes extending the framework to broader biomedical domains and improving evaluation with real-world, open-ended queries.

This paper presents a personalized diet recommendation system that leverages a graph database and retrieval-augmented generation (RAG) with large language models (LLMs) to provide accurate and tailored nutritional plans. The system focuses on vegetarian Indian dishes and analyzes individual health metrics such as physical activity levels, weight age and height to generate customized diet plans. A Neo4j graph database, built from scraped recipe data, stores dish names, ingredients, and nutritional values, enabling precise information retrieval via LLMs. Additionally, the system provides nutritional details and overall health assessments of the dishes. To enhance overall health, the system also suggests basic workout plans based on users' physical metrics and dietary recommendations. This approach aims to provide a low-cost, accessible alternative to traditional diet consultations while mitigating hallucinations in AI outputs through RAG and structured data.

Nutritional recognition has become a critical component in modern healthcare, which is used in personalized dietary management and chronic disease prevention. Traditional approaches, ranging from manual analysis to deep learning methods, face limitations in scalability, real-world adaptability. The advent of large language models (LLMs) shows transformative potential to address these challenges, which leverage multimodal integration and cross-modal attention mechanisms. This paper systematically reviews the applications of LLMs in nutritional recognition. Key innovations such as nutrition-specific tokenization and vision-language alignment demonstrate significant improvements in accuracy and practical utility. However, LLMs face challenges, including dataset dependency, computational inefficiency, and Hallucination. By synthesizing recent advancements, this paper also points out the future directions.

… Knowledge Graphs with Retrieval-Augmented Generation (RAG), we enhance biomedical knowledge … for advancing personalized nutrition and multi-domain biomedical applications. …

Grounding diagnostic reasoning in structured knowledge is critical for applying AI to Traditional Chinese Medicine (TCM), where accurate syndrome differentiation remains highly dependent on physician expertise and hindered by the unstructured, often ambiguous nature of classical texts [1]. This dependency introduces substantial variability in diagnostic outcomes and poses a critical barrier to standardization and AI-assisted clinical deployment. Here, we present TCM-RAG, a Retrieval-Augmented Generation (RAG) system grounded in a domain-specific, structured knowledge graph (KG) explicitly designed to model the core clinical reasoning pathway for TCM spleen-stomach disorders. Our approach makes three key contributions: (1) We design and implement a KG schema that captures the essential "Syndrome → Treatment Principle → Formula" inferential chain; (2) We introduce a path-based tem-plate method to automatically generate a large-scale, clinically consistent diagnostic dataset from classical texts, overcoming data scarcity and annotation inconsistency [2]; and (3) We demonstrate through rigorous quantitative evaluation that our RAG framework—without any model fine-tuning—significantly outperforms zero-shot LLM baselines by leveraging structured retrieval for context-aware, clinically grounded generation. Experimental results on a 500-case benchmark show our system achieves 51.7% recall for syndrome differentiation and 42.5% end-to-end reasoning path completeness (Syndrome + Formula), substantially exceeding baseline performance (absolute gains: +9.4 pp / +11.3 pp; relative gains: +22.2% for Syndrome Recall, +36.2% for Path Completeness). Error analysis confirms our system’s outputs are clinically coherent, logically consistent, and free from hallucinated syndromes or contradictory treatments. These findings establish the feasibility and superiority of structured knowledge-guided LLMs for complex, logic-intensive medical domains like TCM, offering a scalable, interpretable pathway toward trustworthy AI-assisted diagnosis.

: Persona is not a static demographic label (“25-year-old student”) but an adaptive, evolving representation of the user. We propose a dynamic persona model that acts as a living mirror of the user, constantly adapting to their mental state, context, and behavior, and feeding that into food suggestions that support emotional well-being. The model combines relatively stable attributes (e.g., dietary preferences, allergies) with time-sensitive states (e.g., stress). We operationalize this through state charts and ontologies that link mental states with nutrition recommendations grounded in nutritional psychiatry. The resulting hybrid pipeline integrates ontological reasoning with adaptive learning to continuously update the state and recommend context-appropriate foods aimed at stabilizing or improving well-being. A proof-of-feasibility prototype demonstrates how state transitions can trigger timely adjustments in food suggestions without compromising nutritional adequacy and user constraints. This work positions dynamic personas as contextual twins that evolve with the user, enabling explainable and responsive food recommendations. The work also establishes the feasibility for integrating multimodal data streams from smart devices (e.g., wearables, smart kitchen tools, smart plates, smart fridges) to capture daily fluctuations relevant to mental health and link them semantically to food-related ontologies.

… risks, particularly in high-volume regions. To address this, we developed a novel risk assessment … The framework implements LLM-enhanced data governance, a Delphi-AHP dual-…

,

• Automatic extraction of food safety hazards from literature can be a promising tool. • Large language models can be applied for information extraction from literature. • Extra training of large language models is not necessary; prompting is sufficient. • A prompt breaking the task down into smaller steps performs best. • Chemical hazards were successfully captured for dairy, maize and salmon. The number of scientific articles published in the domain of food safety has consistently been increasing over the last few decades. It has therefore become unfeasible for food safety experts to read all relevant literature related to food safety and the occurrence of hazards in the food chain. However, it is important that food safety experts are aware of the newest findings and can access this information in an easy and concise way. In this study, an approach is presented to automate the extraction of chemical hazards from the scientific literature through large language models. The large language model was used out-of-the-box and applied on scientific abstracts; no extra training of the models or a large computing cluster was required. Three different styles of prompting the model were tested to assess which was the most optimal for the task at hand. The prompts were optimized with two validation foods (leafy greens and shellfish) and the final performance of the best prompt was evaluated using three test foods (dairy, maize and salmon). The specific wording of the prompt was found to have a considerable effect on the results. A prompt breaking the task down into smaller steps performed best overall. This prompt reached an average accuracy of 93 % and contained many chemical contaminants already included in food monitoring programs, validating the successful retrieval of relevant hazards for the food safety domain. The results showcase how valuable large language models can be for the task of automatic information extraction from the scientific literature.

… As the food industry continues to evolves and global trade expands, food safety challenges … applications including quality assessment and upstream agriculture risks. Current challenges …

… safety-critical entities and relationships from unstructured sources. This review presents the latest advances in KG-LLM integration for food safety… responses to proactive risk assessment. …

such as PubChem, ChEMBL, ACToR, and Tox21/ToxCast, the models can be fine-tuned to better extract and analyze chemical data. In addition, predictive toxicology models like the DeepTox have demonstrated significant augmentation in the process of risk characterization in recent times. Interestingly, the European Food Safety Authority (EFSA) has investigated the potential of LLMs in toxicology within the AI4NAMS project by evaluating their performance in handling data on Bisphenol A (BPA). The study benchmarked the performance of a baseline Generative Pre-trained Transformer (GPT) model against a fine-tuned model, with the latter proving to be more effective in extracting and consolidating relevant toxicological data

… Model (LLM)-generated summaries of food safety documents, plus food safety related … food safety data, scientific articles, and risk assessments to support the development of new …

Background Nutritional management for patients with diabetes in China is a significant challenge due to the low supply of registered clinical dietitians. To address this, an artificial intelligence (AI)–based nutritionist program that uses advanced language and image recognition models was created. This program can identify ingredients from images of a patient’s meal and offer nutritional guidance and dietary recommendations. Objective The primary objective of this study is to evaluate the competence of the models that support this program. Methods The potential of an AI nutritionist program for patients with type 2 diabetes mellitus (T2DM) was evaluated through a multistep process. First, a survey was conducted among patients with T2DM and endocrinologists to identify knowledge gaps in dietary practices. ChatGPT and GPT 4.0 were then tested through the Chinese Registered Dietitian Examination to assess their proficiency in providing evidence-based dietary advice. ChatGPT’s responses to common questions about medical nutrition therapy were compared with expert responses by professional dietitians to evaluate its proficiency. The model’s food recommendations were scrutinized for consistency with expert advice. A deep learning–based image recognition model was developed for food identification at the ingredient level, and its performance was compared with existing models. Finally, a user-friendly app was developed, integrating the capabilities of language and image recognition models to potentially improve care for patients with T2DM. Results Most patients (182/206, 88.4%) demanded more immediate and comprehensive nutritional management and education. Both ChatGPT and GPT 4.0 passed the Chinese Registered Dietitian examination. ChatGPT’s food recommendations were mainly in line with best practices, except for certain foods like root vegetables and dry beans. Professional dietitians’ reviews of ChatGPT’s responses to common questions were largely positive, with 162 out of 168 providing favorable reviews. The multilabel image recognition model evaluation showed that the Dino V2 model achieved an average F1 score of 0.825, indicating high accuracy in recognizing ingredients. Conclusions The model evaluations were promising. The AI-based nutritionist program is now ready for a supervised pilot study.

Obesity is currently classified as a global epidemic that is a primary cause of cardiovascular and other diseases. According to the World Bank, obesity related factors are the second largest cause of mortality in the modern world. Among the primary causes of obesity, environmental factors like fast food consumption and sedentary lifestyle have been identified and the role of health coaches in promoting a healthy lifestyle is now medically established. Behavior modification therapy through guided and encouraging health coaching substantially aids in reducing obesity. Yet, due to costliness and logistical challenges, access to a human health coach is limited to the developed world and the physically agile population. In this project, I create DigiMate, a Web-based application that harnesses the power of large language models to provide digital health coaching customized to the user health requirements. Using a mathematical dynamic general equilibrium model based on lifetime utility maximization, I first trace how behavior modification therapy works through lowering the lifetime health cost of obesity. DigiMate works by utilizing ChatGPT to generate customized health messages based on user inputs to aid behavior modification. Randomized control trial on 63 seniors shows significantly reduced fast food consumption and increased exercise time for continuous DigiMate use for a period of 3 months for the target group as compared to the control group without DigiMate access.

Background/Objectives: Weight management is directly linked to cancer recurrence and survival, but unfortunately, nutritional oncology counseling is not typically covered by insurance, creating a disparity for patients without nutritional education and food access. Novel ways of imparting personalized nutrition advice are needed to address this issue. Large language models (LLMs) offer a promising path toward tailoring dietary advice to individual patients. This study aimed to assess the capacity of LLMs to offer personalized dietary advice to patients with breast cancer. Methods: Thirty-one prompt templates were designed to evaluate dietary recommendations generated by ChatGPT and Gemini with variations within eight categorical variables: cancer stage, comorbidity, location, culture, age, dietary guideline, budget, and store. Seven prompts were selected for four board-certified oncology dietitians to also respond to. Responses were evaluated based on nutritional content and qualitative observations. A quantitative comparison of the calories and macronutrients of the LLM- and dietitian-generated meal plans via the Acceptable Macronutrient Distribution Ranges and United States Department of Agriculture’s estimated calorie needs was performed. Conclusions: The LLMs generated personalized grocery lists and meal plans adapting to location, culture, and budget but not age, disease stage, comorbidities, or dietary guidelines. Gemini provided more comprehensive responses, including visuals and specific prices. While the dietitian-generated diets offered more adherent total daily calorie contents to the United States Department of Agriculture’s estimated calorie needs, ChatGPT and Gemini offered more adherent macronutrient ratios to the Acceptable Macronutrient Distribution Range. Overall, the meal plans were not significantly different between the LLMs and dietitians. LLMs can provide personalized dietary advice to cancer patients who may lack access to this care. Grocery lists and meal plans generated by LLMs are applicable to patients with variable food access, socioeconomic means, and cultural preferences and can be a tool to increase health equity.

BACKGROUND Prominent large language models, such as OpenAI's ChatGPT, have shown promising implementation in the field of nutrition. Special care should be taken when using ChatGPT to prescribe protein restricted diets for kidney impaired patients. The objective of the current study is to simulate a chronic kidney disease (CKD) patient and evaluate the capabilities of ChatGPT in the context of dietary prescription, with a focus on protein contents of the diet. METHOD We simulated a scenario involving a CKD patient and replicated a clinical counseling session that covered general dietary principles, dietary assessment, energy and protein recommendation, dietary prescription, and diet customization based on dietary culture. To confirm the results derived from our qualitative observations, ten colleagues were recruited and provided with identical dietary prescription prompts to run the process again. The actual energy and protein levels of the given meal plans were recorded and the difference from the targets were compared. RESULTS ChatGPT provides general principles overall aligning with best practices. The recommendations for energy and protein requirements of CKD patients were tailored and satisfactory. It failed to prescribe a reliable diet based on the target energy and protein requirements. For the quantitative analysis, the prescribed energy levels were generally lower than the targets, ranging from -28.9% to -17.0% and protein contents were tremendously higher than the targets, ranging from 157% to 59.3%. CONCLUSION ChatGPT is competent in offering generic dietary advice, giving satisfactory nutrients recommendations and adapting cuisines to different cultures but failed to prescribe nutritionally accurate dietary plans for CKD patients. At present, patients with strict protein and other particular nutrient restrictions are not recommended to rely on the dietary plans prescribed from ChatGPT to avoid potential health risks.

Background: Chronic kidney disease (CKD) requires strict dietary management tailored to disease stage and individual needs. Recent advances in artificial intelligence (AI) have introduced chatbot-based tools capable of generating dietary recommendations. However, their accuracy, personalization, and practical applicability in clinical nutrition remain largely unvalidated, particularly in non-Western settings. Methods: Simulated patient profiles representing each CKD stage were developed and used to prompt GPT-4 (OpenAI), Gemini (Google), and Copilot (Microsoft) with the same request for meal planning. AI-generated diets were evaluated by three physicians using a 5-point Likert scale across three criteria: personalization, consistency with guidelines, practicality, and availability. Descriptive statistics, Kruskal–Wallis tests, and Dunn’s post hoc tests were performed to compare model performance. Nutritional analysis of four meal plans (Initial, GPT-4, Gemini, and Copilot) was conducted using both GPT-4 estimates and manual calculations validated against clinical dietary sources. Results: Scores for personalization and consistency were significantly higher for Gemini and GPT-4 compared with Copilot, with no significant differences between Gemini and GPT-4 (p = 0.0001 and p = 0.0002, respectively). Practicality showed marginal significance, with GPT-4 slightly outperforming Gemini (p = 0.0476). Nutritional component analysis revealed discrepancies between GPT-4’s internal estimations and manual values, with occasional deviations from clinical guidelines, most notably for sodium and potassium, and moderate overestimation for phosphorus. Conclusions: While AI chatbots show promise in delivering dietary guidance for CKD patients, with Gemini demonstrating the strongest performance, further development, clinical validation, and testing with real patient data are needed before AI-driven tools can be fully integrated into patient-centered CKD nutritional care.

In recent years, an increasing number of individuals have turned to traditional Chinese medicine diet therapy as a means to nourish their bodies and mitigate diseases. With the advent of the big data era, knowledge graphs, as powerful analysis tools, can provide more accurate and personalized dietary advice for diet therapy. However, most of the current diet therapy knowledge graphs have imperfections. To address this issue, we construct a diet therapy knowledge graph by utilizing textual data and professional books provided by the Academy of Traditional Chinese Medicine, from which we extract entities and relations. Building upon this foundation, we introduce a text representation technique predicated on contrastive learning, designed to augment the semantic richness of the knowledge graph and enhance the completion of the diet therapy knowledge graph. By conducting experiments on the diet therapy knowledge graph and public datasets, the results show that our method can capture the semantic information in the knowledge graph more efficiently compared to traditional methods. This provides new possibilities for research and practice in the field of traditional Chinese medicine diet therapy. This research opens new avenues for leveraging big data analysis in traditional Chinese medicine diet therapy.

ABSTRACT Objectives Large Language Models (LLM) like ChatGPT and Gemini have potential in nutrition applications, but recent studies suggest they provide inaccurate dietary advice. The aim of this study was to evaluate the most commonly used LLMs, ChatGPT and Gemini, for dietary recommendations for patients with irritable bowel syndrome (IBS). Methods Various tools were used to assess the responses of LLMs in this study. The Guideline Compliance Score was created using IBS guidelines. The quality of the responses provided by LLMs was assessed using The Global Quality Score (GQS) and Completeness, Lack of Misinformation, Evidence, Appropriateness, Relevance (CLEAR) tool. Understandability and actionability were assessed using the Patient Education Materials Assessment Tool (PEMAT). The readability of ChatGPT and Gemini's responses was evaluated using Flesch Reading Ease (FRE) and Flesch Kincaid Grade Level (FKGL). Results This study found that most responses from ChatGPT (70%) and Gemini (57.5%) were compliant with the guidelines, but there was no significant difference in guideline compliance, quality, understandability, actionability, or readability scores ( p &gt; 0.05). The CLEAR tool showed a moderate positive correlation with PEMAT actionability ( r = 0.467, p = 0.038) and understandability ( r = 0.568, p = 0.009), a strong positive correlation with GQS ( r = 0.611, p = 0.004). In addition, FRE and FKGL had a strong negative correlation ( r = −0.784, p &lt; 0.001), while the Guideline Compliance Score showed a moderate negative correlation with FRE ( r = −0.537, p = 0.015). Conclusions The study emphasizes the need for further model improvements before relying solely on LLMs in clinical nutrition practice, emphasizing the importance of dietitians' recommendations and the collaboration between AI models and healthcare teams.

Objective Although artificial intelligence (AI)-based nutrition recommendations are becoming increasingly common among the public, the accuracy and reliability of diets produced especially for adolescents in the growth and development period are not sufficiently known. This study aimed to evaluate the clinical validity of AI by comparing the nutritional content of diets generated by different AI models with dietitian reference plans. Methods A total of 60 three-day diet plans were generated in two sessions by five AI models (ChatGPT-4o, Gemini 2.5 Pro, Claude 4.1, Bing Chat-5GPT, and Perplexity) for four standardized adolescent profiles in this cross-sectional and comparative study. A dietitian reference plan was prepared for each profile. Energy and macro-micronutrients were analyzed with BeBiS. Comparisons were evaluated with single-sample t-test, Cohen’s d, and Bland–Altman fit analyses. Results AI models tended to systematically undercalculate energy (bias: +695 kcal), protein (+19.9 g), lipid (+15.8 g), and carbohydrate (+114.6 g). In macronutrient percentages, protein (21.5–23.7%) and lipid (41.5–44.5%) ratios were above the recommended adolescent guidelines, while carbohydrate ratios (32.4–36.3%) were significantly below. Significant variation was observed between models in micronutrient contents, and no model showed consistent proximity to the dietitian across all nutrients. Conclusion AI models have exhibited clinically significant deviations in diet plans for adolescents at both macro and micro levels. The findings indicate that AI-based dietary recommendations are not appropriate to use without professional supervision, emphasizing the need for model improvements for more reliable data generation in this area.

The chatbot Chat Generative Pretrained Transformer (ChatGPT) is becoming increasingly popular among patients for searching health-related information. Prior studies have raised concerns regarding accuracy in offering nutritional advice. We investigated in November 2023 ChatGPT’s potential as a tool for providing nutritional guidance in relation to different non-communicable diseases (NCDs). First, the dietary advice given by ChatGPT (version 3.5) for various NCDs was compared with guidelines; then, the chatbot’s capacity to manage a complex case with several diseases was investigated. A panel of nutrition experts assessed ChatGPT’s responses. Overall, ChatGPT offered clear advice, with appropriateness of responses ranging from 55.5% (sarcopenia) to 73.3% (NAFLD). Only two recommendations (one for obesity, one for non-alcoholic-fatty-liver disease) contradicted guidelines. A single suggestion for T2DM was found to be “unsupported”, while many recommendations for various NCDs were deemed to be “not fully matched” to the guidelines despite not directly contradicting them. However, when the chatbot handled overlapping conditions, limitations emerged, resulting in some contradictory or inappropriate advice. In conclusion, although ChatGPT exhibited a reasonable accuracy in providing general dietary advice for NCDs, its efficacy decreased in complex situations necessitating customized strategies; therefore, the chatbot is currently unable to replace a healthcare professional’s consultation.

OBJECTIVE Large Language Models (LLMs) have emerged as powerful tools with significant potential for quickly accessing information in the nutrition and health, as in many fields. Retrieval augmented generation (RAG) has been included among artificial intelligence (AI) powered chatbot structures as a framework developed to increase the accuracy and ability of LLMs. This study aimed to evaluate the accuracy of LLMs (GPT4, Gemini, and Llama) and RAG in determining dietary principles in chronic kidney disease. DESIGN AND METHODS The nutrition guideline published by the National Kidney Foundation in 2020 was used as an external information source in developed RAG model. Answers were obtained using 12 medical nutritional therapy prompts for CKD by four chatbots. The accuracy of the 48 answers generated by the chatbots was evaluated with a 5-point Likert scale. RESULTS The results showed that Gemini and RAG had the highest accuracy scores (median:4.0), followed by GPT4 (median: 2.5) and Llama (median: 1.5), respectively. When the accuracy scores were examined between the two chatbots, a significant difference was detected between all groups except Gemini and RAG. CONCLUSION These chatbots produced both completely correct answers and false information with potentially harmful clinical outcomes. Customization of LLMs in specific areas such as nutrition or the development of a nutrition-specific RAG framework by improving LLM structures with current guidelines and articles may be an important strategy to increase the accuracy of AI powered chatbots.

BACKGROUND Our previous research indicated that ChatGPT-3.5 was inadequate in generating nutritionally accurate dietary plans for patients with chronic kidney disease (CKD). Given the subsequent release of ChatGPT-40 and the emergence of DeepSeek, the aim of this study is to assess and compare the CKD dietary plans generated by ChatGPT-3.5, ChatGPT-4.0, and DeepSeek in terms of both nutritional accuracy and the Dietary Inflammatory Index (DII). METHODS Standardized dietary prescription prompts with a fixed energy target of 1800 kcal and five protein levels (30 g, 40 g, 50 g, 60 g, and 70 g) were designed to represent different stages of CKD. Twenty participants were recruited and instructed to input these pre-established prompts into the three LLMs to generate renal diets. The actual energy and protein contents of the resulting meal plans were recorded and compared across the three models. The DII values were calculated based on the USDA database. RESULTS A total of 300 meal plans were generated. All three LLMs failed to meet the 1800 kcal energy target, with values ranging from 1165 to 1548 kcal (P < 0.001). DeepSeek generated plans that were closest to the target energy and protein levels. ChatGPT-4.0 generated the most anti-inflammatory plans (DII = -1.487335), followed by ChatGPT-3.5 (DII = -0.9163290); whereas DeepSeek (DII = 2.2502160) produced pro-inflammatory plans. CONCLUSION This study confirmed the limitations of LLMs in generating nutritionally accurate dietary plans for patients with CKD. DeepSeek outperformed the ChatGPT models in nutritional accuracy, whereas the ChatGPT models generated more anti-inflammatory diets. As AI evolves, human oversight remains crucial in ensuring the safety and efficacy of AI-generated dietary plans.

OBJECTIVE Inpatients undergoing stroke rehabilitation experience high malnutrition rates, requiring strict dietary management. However, manual and time-pressured dietary provision can cause errors in diet composition, highlighting the need for innovation. Therefore, we aimed to evaluate whether GPT-4o can accurately identify dietary errors in hospital-based stroke rehabilitation menus, analyze differences in AI vs. expert rationale for decisions, and explore AI's potential role in clinical workflows through a structured collaboration framework. METHODS A TRIPOD-compliant validation study analyzing 264 hospital-based menus designed for stroke rehabilitation inpatients requiring specialized diets (e.g., dysphagia, diabetes). GPT-4o's dietary compliance classifications were assessed using a structured 0-error, 1-error, and 2+ error framework, with expert dietitians as ground-truth in a rehabilitation hospital nutrition department, where expert dietitians selected menus from existing clinical practices for inpatients on specialized diets. AI-expert agreement, overall accuracy, sensitivity, and specificity in dietary error classification were assessed. AI vs. expert justifications were analyzed thematically to identify differences in decision rationale. Cohen's Kappa (95% CI) measured inter-rater reliability. Overall accuracy, sensitivity, and specificity were calculated using a 3 × 3 confusion matrix, comparing AI classifications (0-error, 1-error, 2+ error) to the expert-labeled ground truth. Thematic analysis categorized AI vs. expert justifications for flagged dietary errors. RESULTS Out of 264 menus (1,000+ food items), 26 (9.8%) had discrepancies. Among these, 57.7% (15 cases) were PAS-based dysphagia diets, followed by diabetic (19.2%, 5 cases) and allergen-related (15.4%, 4 cases) diets. The remaining two cases involved low-sodium and low-fat diets. Cohen's Kappa: 0.892 (95% CI: 0.845-0.939, p < 0.001). 0-errors: Sensitivity 94.3%, specificity 100%; 1-error: Sensitivity 86.2%, specificity 96.6%; 2+-errors: Sensitivity 97.8%, specificity 92.6%. Thematic analysis revealed GPT-4o followed strict rule-based interpretations, whereas dietitians incorporated patient tolerance and food preparation considerations. CONCLUSION GPT-4o demonstrated high accuracy but over-flagged violations, supporting its role as a prescreening tool with expert collaboration.