LB076 - CAN AI READ A FOOD LABEL? A MULTIMODAL EVALUATION OF LARGE LANGUAGE MODELS CLASSIFYING PACKAGED FOODS BY NUTRI-SCORE, NOVA, HEALTH STAR RATING AND UK TRAFFIC-LIGHT SYSTEMS FOR OBESITY-RELATED DIETARY COUNSELLING
LB076
CAN AI READ A FOOD LABEL? A MULTIMODAL EVALUATION OF LARGE LANGUAGE MODELS CLASSIFYING PACKAGED FOODS BY NUTRI-SCORE, NOVA, HEALTH STAR RATING AND UK TRAFFIC-LIGHT SYSTEMS FOR OBESITY-RELATED DIETARY COUNSELLING
D. Düman1,*, S. N. Kılıç1, B. Temur1, I. Güler1, Y. Aykemat1, N. İnanç1
1Nutrition and Dietetics, Nuh Naci Yazgan University, Kayseri, Türkiye
Rationale: High intake of ultra-processed foods and poor diet quality are major modifiable contributors to obesity and metabolic syndrome. Front-of-pack food-processing classification systems, including Nutri-Score, NOVA, Health Star Rating (HSR) and UK traffic-light scheme, support healthier food choices but can be time-consuming to apply. The ability of large language models to accurately reproduce these classifications has not been evaluated.
Methods: In this cross-sectional diagnostic-accuracy study, packaged foods from the Turkish retail market will be sampled across obesity-relevant categories, including sugar-sweetened beverages, snacks, ready meals, breakfast cereals, and processed meats. Products will be classified using four established systems. Reference standards will comprise official nutrient-profiling algorithms and independent dual-expert NOVA coding. Multimodal LLMs will be evaluated using structured text inputs and product-label photographs. Primary outcomes will include agreement with reference classifications (weighted Cohen’s κ), while secondary outcomes will assess classification accuracy, NOVA-4 detection performance, modality effects, between-model differences,and test–retest reliability
Results: Data collection and analyses suggest variability in classification performance across LLMs and food-classification systems. Differences were observed between text- and image-based inputs, with agreement levels varying according to the classification framework applied.
Conclusion: Initial findings indicate that LLM-based food classification may be feasible, although performance appears to depend on both the model and the classification system used. Further analyses will clarify the potential role of AI-assisted approaches in dietary assessment and obesity-related nutrition counselling.
Disclosure of Interest: None declared