Uso da inteligência artificial para estimativa dietética com base em imagens de alimentos

Abstract

Introduction: Artificial intelligence has stood out as a promising tool in the field of nutrition, especially in studies investigating its application in assessing food consumption through images. Objective: To evaluate the correlation and agreement between the analysis of the nutrient composition of food images in ChatGPT and nutrition software. Materials and methods: The research was conducted using images of some foods belonging to the groups of cereals and tubers, pulses, meats, vegetables, and greens available in photographic atlas. Estimates from the analysis of images by ChatGPT were compared with the evaluation by software. Results: ChatGPT correctly identified most foods, but tended to overestimate the weight and total energetic value of medium and large portions of foods, especially those with higher energy density, and to overestimate the carbohydrate and protein content of foods with low energy density. Agreement between the methods generally ranged from poor to moderate. The presentation of household measures in food images (spoons, ladles, and skimmers) resulted in lower standard deviations, leading to greater consistency in estimates. Conclusion: ChatGPT has significant potential as a complementary tool for image-based dietary assessment, especially for rapid screening and preliminary analysis. However, image standardization and model improvement for regional foods are necessary, as well as the application of ethical and privacy issues. Thus, future studies with larger samples are needed to consolidate its applicability in nutritionist practice.