Abstract Background and Aims The rising diversity of food preferences and the desire to provide better personalized care provide challenges to renal dietitians working in dialysis clinics. To address this situation, we explored the use of a Large Language Model (LLM), specifically, ChatGPT using the GPT-4 model (openai.com), to support nutritional advice given to dialysis patients. Method We applied a prompt that allows us to consider various factors that determine personalized dietary recommendations, menu options, and nutritional analysis. The prompt integrates several domains, including patient demographics (age, sex, race, and ethnicity) and food preferences (vegetarian, Indian, Latin American, Mediterranean, East Asian), laboratory data (levels of serum phosphate, potassium, albumin, calcium, and normalized protein catabolic rate), and clinical characteristics (height, BMI, comorbidities, and dialysis modality). Twenty virtual patients were created using Monte Carlo simulation in the statistical software R; we provided the characteristics of one randomly selected patient to ChatGPT (Fig. 1). The resulting daily recipe recommendations, cooking instructions, and nutritional analyses were reviewed and rated on a five-point Likert scale by an experienced renal dietitian. In addition, we entered the ChatGPT-generated recipes into an USDA-approved nutrient analysis software (ESHA research, Salem, OR, USA) for comparison. A pre-defined kidney-friendly recipe from the website of a large dialysis organization (Fresenius Medical Care (FMC), Waltham, MA, USA) was included for comparison as well. We were able to narrow down and personalize requests in a short dialogue with the bot, for example: “John is a hemodialysis patient. Age, years: 63.58; Height, cm: 181.52; BMI: 37.27; Serum albumin, g/dL: 3.08; Serum phosphate, mg/dL: 4.39; Serum potassium, mmol/L: 5.08; Total serum calcium, mg/dL: 8.63; nPCR, g/kg/day: 0.709; Sex: Male; Race/Ethnicity: Black; DM: No; CHF: No; Diet Restriction: No restriction; Food preferences: Latin American; Dialysis modality: HD; Budget: around $25. Can you create a one-day sample menu for this hemodialysis patient with recipes and nutritional analyses?” Results A daily menu comprising five recipes was created. The renal dietitian rated the recipes (2.9 on the 5-point Likert scale) and cooking instructions (4.8) as moderately satisfactory, but the nutritional analysis (2) was questionable. While ChatGPT's carbohydrate content estimation was relatively accurate, it underestimated calories by 36%, protein by 28%, and fat, phosphorus, potassium, and sodium by around 50% (Fig. 2A). Discrepancies observed with potassium may be partly explained by the analyses of raw vs. cooked ingredients (which may be altered by cooking techniques), unspecified portion sizes, and limited data for ChatGPT training (e.g., potassium is not a required nutrition label component). Interestingly, these discrepancies are much smaller with recipes garnered from the FMC website (Fig. 2B). The percentage differences between USDA's (=100%) and ChatGPT's estimations of recipes from different sources were shown in Fig. 2C. In addition, we explored the use of ChatGPT to create personalized dietary recommendations/menus in different languages, such as Mandarin. The translations, as judged by native speakers, were reliable. Conclusion While LLMs such as ChatGPT hold promise to provide personalized nutritional guidance to diverse populations of dialysis patients, we think that there is substantial room for improvement. Our findings underscore the importance of a critical approach and the need to assess qualitatively and quantitatively the output created by LLMs, especially regarding medical use cases.
Read full abstract