PP376 - CAN ARTIFICIAL INTELLIGENCE SUSTAIN SUSTAINABLE DIETS? A LONGITUDINAL ASSESSMENT OF MENU QUALITY AND VARIABILITY?

PP376

CAN ARTIFICIAL INTELLIGENCE SUSTAIN SUSTAINABLE DIETS? A LONGITUDINAL ASSESSMENT OF MENU QUALITY AND VARIABILITY?

H. Küçükkatırcı Baykan¹, S. Calapkorur^2,*

¹Nutrition and Dietetic, Kapadokya University School of Health Sciences Department of Nutrition and Dietetics, Nevşehir, ²Nutrition and Dietetic, Erciyes University Health Science Faculty, Kayseri, Türkiye

Rationale: Environmental sustainability and artificial intelligence (AI) are key drivers in modern nutrition. However, integrating Large Language Models (LLMs) into medical nutrition therapy raises concerns regarding algorithmic reliability. This in silico longitudinal study evaluated the temporal consistency and output fluctuations of ChatGPT-4 in generating menus for seven sustainable diets (Mediterranean, Barilla, DASH, Vegetarian, Vegan, New Nordic, Flexitarian).

Methods: Using synthetic patient profiles, sequential menus were generated at baseline, week 1, and month 1. Performance was quantified via the MedQ-Sus-Tr index. Statistical reliability was assessed using Intraclass Correlation Coefficient (ICC) for temporal consistency and Coefficient of Variation (CV%) for algorithmic stability and prompt drift.

Results: The AI produced superficial sustainability profiles, failing to distinguish between diets with diverse ecological foundations. ICC analyses showed "poor" consistency (ICC < 0.50), indicating severe instability across all models. While the Vegan model was initially stable due to simple constraints, it reached a 24.04% fluctuation by month 1 due to prompt drift. The New Nordic model exhibited high variance throughout. In the vegetarian model, score improvements were misleading, as the algorithm became trapped in limited food cycles, significantly reducing dietary diversity.

Image:

Conclusion: Day-to-day algorithmic inconsistency risks patient adherence by increasing decision fatigue. LLMs currently lack the clinical reliability for autonomous management of complex dietary patterns. They should function solely as auxiliary tools under the professional supervision of a dietitian (human-in-the-loop) to ensure nutritional safety and diversity.

Disclosure of Interest: None declared