Personalized AI-Powered Food Ingredient Analysis and Dietary Recommendation

ABSTRACT Introduction Interpreting complex food labels and aligning them with individual health needs remains a challenge for consumers. This pilot study evaluated the ability of large language models (LLMs) to analyze structured patient profiles and provide personalized nutritional guidance based on medical history and food ingredient data. Methods Twenty synthetic patient profiles were generated, varying by age, sex, height, weight, and medical history. Each profile was paired with a set of 10 food items (categorized as healthy, moderate, or unhealthy). These inputs were submitted to four LLMs, including ChatGPT, Claude, Gemini, and DeepSeek. The models were instructed to assess the health impact of each food item per patient and classify it as beneficial (green), moderate (yellow), or harmful (red). Results All LLMs accurately flagged high-sugar or ultra-processed foods as harmful and healthy items as beneficial. However, their classification accuracy and depth of reasoning varied: ChatGPT (72.5%), Claude (67%), Gemini (63.5%), and DeepSeek (63%). Only ChatGPT consistently provided both accurate classifications and personalized, relevant alternatives. Other models lacked consistency, specificity, or justification in their responses. Conclusion ChatGPT demonstrated the highest accuracy and interpretability, suggesting its potential utility in delivering personalized dietary recommendations based on individual health data. Integrating advanced LLMs into digital health tools could enhance nutrition guidance, especially for individuals with diet-sensitive conditions such as diabetes or obesity.