食品图像营养成分估计的3种大型语言模型性能评价

IF 3.2 Q2 NUTRITION & DIETETICS

Current Developments in Nutrition Pub Date : 2025-10-01 DOI:10.1016/j.cdnut.2025.107556

Jonatan Fridolfsson , Emma Sjöberg , Meri Thiwång , Stefan Pettersson

{"title":"食品图像营养成分估计的3种大型语言模型性能评价","authors":"Jonatan Fridolfsson , Emma Sjöberg , Meri Thiwång , Stefan Pettersson","doi":"10.1016/j.cdnut.2025.107556","DOIUrl":null,"url":null,"abstract":"<div><h3>Background</h3><div>Traditional dietary assessment methods face limitations including recall bias, participant burden, and portion size estimation errors. Recent advances in artificial intelligence, particularly multimodal large language models (LLMs), offer potential solutions for automated nutritional analysis from food images.</div></div><div><h3>Objectives</h3><div>This study aims to evaluate and compare the performance of 3 leading LLMs (ChatGPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro) in estimating food weight, energy content, and macronutrient composition from standardized food photographs.</div></div><div><h3>Methods</h3><div>We analyzed 52 standardized food photographs including individual food components (<em>n</em> = 16) and complete meals (<em>n</em> = 36) in 3 portion sizes (small, medium, large). Each model received identical prompts to identify food components and estimate nutritional content using visible cutlery and plates as size references. Model estimates were compared against reference values obtained through direct weighing and nutritional database analysis (Dietist NET). Performance metrics included mean absolute percentage error (MAPE), Pearson correlations, and systematic bias analysis using Bland–Altman plots.</div></div><div><h3>Results</h3><div>ChatGPT and Claude demonstrated similar accuracy with MAPE values of 36.3% and 37.3% for weight estimation, and 35.8% for energy estimation. Gemini showed substantially higher errors across all nutrients (MAPE 64.2%–109.9%). Correlations between model estimates and reference values ranged from 0.65 to 0.81 for ChatGPT and Claude, compared with 0.58–0.73 for Gemini. All models exhibited systematic underestimation that increased with portion size, with bias slopes ranging from –0.23 to –0.50.</div></div><div><h3>Conclusions</h3><div>ChatGPT and Claude achieved accuracy levels comparable with traditional self-reported dietary assessment methods but without associated user burden, suggesting potential utility as dietary monitoring tools. However, systematic underestimation of large portions and high variability in macronutrient estimation indicate these general-purpose LLMs are not yet suitable for precise dietary assessment in clinical or athletic populations where accurate quantification is critical.</div></div>","PeriodicalId":10756,"journal":{"name":"Current Developments in Nutrition","volume":"9 10","pages":"Article 107556"},"PeriodicalIF":3.2000,"publicationDate":"2025-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Performance Evaluation of 3 Large Language Models for Nutritional Content Estimation from Food Images\",\"authors\":\"Jonatan Fridolfsson , Emma Sjöberg , Meri Thiwång , Stefan Pettersson\",\"doi\":\"10.1016/j.cdnut.2025.107556\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><h3>Background</h3><div>Traditional dietary assessment methods face limitations including recall bias, participant burden, and portion size estimation errors. Recent advances in artificial intelligence, particularly multimodal large language models (LLMs), offer potential solutions for automated nutritional analysis from food images.</div></div><div><h3>Objectives</h3><div>This study aims to evaluate and compare the performance of 3 leading LLMs (ChatGPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro) in estimating food weight, energy content, and macronutrient composition from standardized food photographs.</div></div><div><h3>Methods</h3><div>We analyzed 52 standardized food photographs including individual food components (<em>n</em> = 16) and complete meals (<em>n</em> = 36) in 3 portion sizes (small, medium, large). Each model received identical prompts to identify food components and estimate nutritional content using visible cutlery and plates as size references. Model estimates were compared against reference values obtained through direct weighing and nutritional database analysis (Dietist NET). Performance metrics included mean absolute percentage error (MAPE), Pearson correlations, and systematic bias analysis using Bland–Altman plots.</div></div><div><h3>Results</h3><div>ChatGPT and Claude demonstrated similar accuracy with MAPE values of 36.3% and 37.3% for weight estimation, and 35.8% for energy estimation. Gemini showed substantially higher errors across all nutrients (MAPE 64.2%–109.9%). Correlations between model estimates and reference values ranged from 0.65 to 0.81 for ChatGPT and Claude, compared with 0.58–0.73 for Gemini. All models exhibited systematic underestimation that increased with portion size, with bias slopes ranging from –0.23 to –0.50.</div></div><div><h3>Conclusions</h3><div>ChatGPT and Claude achieved accuracy levels comparable with traditional self-reported dietary assessment methods but without associated user burden, suggesting potential utility as dietary monitoring tools. However, systematic underestimation of large portions and high variability in macronutrient estimation indicate these general-purpose LLMs are not yet suitable for precise dietary assessment in clinical or athletic populations where accurate quantification is critical.</div></div>\",\"PeriodicalId\":10756,\"journal\":{\"name\":\"Current Developments in Nutrition\",\"volume\":\"9 10\",\"pages\":\"Article 107556\"},\"PeriodicalIF\":3.2000,\"publicationDate\":\"2025-10-01\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Current Developments in Nutrition\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S2475299125030185\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q2\",\"JCRName\":\"NUTRITION & DIETETICS\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Current Developments in Nutrition","FirstCategoryId":"1085","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S2475299125030185","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"NUTRITION & DIETETICS","Score":null,"Total":0}

引用次数: 0

摘要

传统的饮食评估方法存在回忆偏倚、参与者负担和分量估计误差等局限性。人工智能的最新进展，特别是多模态大语言模型（llm），为食品图像的自动营养分析提供了潜在的解决方案。本研究旨在评估和比较3种领先的LLMs （chatgpt - 40、Claude 3.5 Sonnet和Gemini 1.5 Pro）在从标准化食品照片中估计食物重量、能量含量和宏量营养素组成方面的性能。方法对52张标准化食品照片进行分析，包括单个食品成分（n = 16）和全餐（n = 36），分为小、中、大3种份量。每个模型都收到了相同的提示，以识别食物成分和估计营养成分，使用可见的餐具和盘子作为尺寸参考。将模型估计值与通过直接称重和营养数据库分析（Dietist NET）获得的参考值进行比较。性能指标包括平均绝对百分比误差（MAPE）、Pearson相关性和使用Bland-Altman图的系统偏差分析。结果schatgpt和Claude的准确率相近，MAPE对权重估计的准确率分别为36.3%和37.3%，对能量估计的准确率为35.8%。Gemini在所有营养成分上的误差要高得多（MAPE为64.2%-109.9%）。ChatGPT和Claude的模型估计值与参考值之间的相关性为0.65 - 0.81，而Gemini的模型估计值为0.58-0.73。所有模型均表现出系统性低估，其偏倚斜率范围为-0.23 ~ -0.50。结论schatgpt和Claude达到了与传统的自我报告饮食评估方法相当的准确性水平，但没有相关的用户负担，提示作为饮食监测工具的潜在效用。然而，对大份量的系统性低估和大量营养素估计的高度可变性表明，这些通用llm尚不适合用于临床或运动人群的精确饮食评估，因为精确量化是至关重要的。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

查看原文本刊更多论文

Performance Evaluation of 3 Large Language Models for Nutritional Content Estimation from Food Images

Background

Traditional dietary assessment methods face limitations including recall bias, participant burden, and portion size estimation errors. Recent advances in artificial intelligence, particularly multimodal large language models (LLMs), offer potential solutions for automated nutritional analysis from food images.

Objectives

This study aims to evaluate and compare the performance of 3 leading LLMs (ChatGPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro) in estimating food weight, energy content, and macronutrient composition from standardized food photographs.

Methods

We analyzed 52 standardized food photographs including individual food components (n = 16) and complete meals (n = 36) in 3 portion sizes (small, medium, large). Each model received identical prompts to identify food components and estimate nutritional content using visible cutlery and plates as size references. Model estimates were compared against reference values obtained through direct weighing and nutritional database analysis (Dietist NET). Performance metrics included mean absolute percentage error (MAPE), Pearson correlations, and systematic bias analysis using Bland–Altman plots.

Results

ChatGPT and Claude demonstrated similar accuracy with MAPE values of 36.3% and 37.3% for weight estimation, and 35.8% for energy estimation. Gemini showed substantially higher errors across all nutrients (MAPE 64.2%–109.9%). Correlations between model estimates and reference values ranged from 0.65 to 0.81 for ChatGPT and Claude, compared with 0.58–0.73 for Gemini. All models exhibited systematic underestimation that increased with portion size, with bias slopes ranging from –0.23 to –0.50.

Conclusions

ChatGPT and Claude achieved accuracy levels comparable with traditional self-reported dietary assessment methods but without associated user burden, suggesting potential utility as dietary monitoring tools. However, systematic underestimation of large portions and high variability in macronutrient estimation indicate these general-purpose LLMs are not yet suitable for precise dietary assessment in clinical or athletic populations where accurate quantification is critical.

求助全文

通过发布文献求助，成功后即可免费获取论文全文。去求助

来源期刊