#922 Evaluating the nutritional accuracy and practicality of ChatGPT-generated recipes for CKD patients

Abstract

Abstract Background and Aims Adequate nutrition is fundamental to human well-being, particularly for patients with impaired kidney function. To address this, we began leveraging artificial intelligence—specifically, ChatGPT using the GPT-4o model (openai.com)—to assist in the development of personalized dietary recommendations for CKD patients. In previous work by Wang et al. (J Ren Nutr, 2024), the GPT-4 model demonstrated qualitative success in generating dietary recommendations for dialysis patients. However, quantitative analysis revealed notable limitations, and the recommendations had not been translated into actual prepared dishes, leaving their practical feasibility untested. This study aims to bridge that gap by preparing dishes based on the LLM's recommendations and evaluating them across multiple dimensions, including appropriateness for specific clinical profiles, palatability, and nutritional accuracy. Method Twenty virtual patients were created using Monte Carlo simulation; we provided the characteristics of three randomly selected patients diverse food preferences (Western, Chinese, and Mexican) to GPT-4o. GPT-4o generated three recipes for each virtual patient based on patient demographics, food preferences, laboratory data, clinical characteristics, and daily food budget. All nine dishes were made in a professional kitchen and rated based on the dishes’ appearance, taste, flavor, texture, portion size, and overall acceptability. We then entered the GPT-4o's-generated recipes into a USDA-approved nutrient analysis software (ESHA research, Salem, OR, USA) for comparison. For each recipe, individual ingredients, their weight and total quantity were cross referenced in the USDA Database for food amounts (fdc.nal.USDA.gov). Three renal dieticians were consulted to cross reference amounts and weights for each ingredient. A recipe card for each dish was generated following this to yield nutritional values for calories, protein, total fat, carbohydrates, phosphorus, potassium and sodium (Fig. 1). Once the nutritional analysis for each recipe was generated, a comparison was made between GPT-4o's nutrient estimation of calories, protein, total fat, carbohydrates, phosphorus, potassium and sodium and ESHA's nutrient analysis per dish. Results While ChatGPT's carbohydrate content estimation was relatively accurate, it underestimated calories by 36%, protein by 28%, and total fat, phosphorus, potassium, and sodium by around 50%. On the other hand, GPT-4o showed an overall improvement where it underestimated calories by 21% (mean: 78.7%; 95% CI: 50.3% to 107.1%), total fat by 36% (63.7%; CI: 46.1% to 81.2%), protein 24% (76.3%; CI: 57.6% to 95.0%), potassium 24% (mean: 75.7%; 95% CI: 54.1% to 97.3%), and overestimated phosphorus by 16% (115.7%; CI: 65.8% to 165.6%). The carbohydrate and sodium contents were satisfactorily accurate (difference within 10%). (Fig. 2). Scatterplots were also generated to compare GPT-4o-generated recipe values to ESHA values for each nutritional value. For instance, Chinese Udon noodles were an outlier in the sodium plot which was attributed to GPT-4o's recommendation for soy sauce – an ingredient high in sodium (Fig. 3). In terms of appearance, dieticians rated the recipes from GPT-4o between 4–5, taste was rated between 3–5, flavor and texture were rated largely around 4, portion size was rated between 4–5 (Fig. 4). Conclusion GPT-4o is a valuable tool for creating personalized nutrition plans for diverse patient groups. However, we recommend cross-checking the nutritional values of recipes with a validated database, such as ESHA, to ensure accuracy. While GPT-4o serves as an excellent starting point for dietitians, ingredient substitutions should align more closely with patients’ dietary needs. For instance, replacing Udon noodles with a healthier alternative like whole wheat noodles or replacing a low sodium soy sauce with fresh lemon or lime juice, and other healthier alternatives. This study underscores the need for continued refinement in AI-generated meal plans and highlights opportunities for collaboration between AI developers and healthcare professionals.

Ähnliche Arbeiten