Czy sztuczna inteligencja i chatboty mogą planować terapeutyczną dietę ketogenną dla dzieci z epilepsją?
Can Artificial Intelligence Chatbots Plan Therapeutic Ketogenic Diets for Children With Epilepsy?
W skrócie
Badanie pokazuje, że popularne chatboty AI (ChatGPT, Microsoft Copilot, Gemini) nie potrafią poprawnie planować diet ketogennych dla dzieci z epilepsją - nie obliczają prawidłowo składników odżywczych i sugerują niewłaściwe potrawy. Wyniki są niepokojące, bo diet ta jest złożona i wymaga precyzyjnych obliczeń, a błędy mogą być niebezpieczne dla zdrowia dzieci. Badacze zaznaczają, że do czasu poprawy, ai chatboty nie powinny być używane do tego celu bez nadzoru wykwalifikowanych specjalistów.
Oryginalny abstract (angielski)
RATIONALE: Ketogenic Diet Therapy (KDT) is an effective but complex treatment for paediatric drug-resistant epilepsy. Access to trained dietitians limits the global use of KDT. The increasing use of artificial intelligence (AI) chatbots for health and dietary advice presents both opportunities and risks. This exploratory study evaluates the accuracy and feasibility of AI-generated KDT plans for paediatric epilepsy. METHODS: We conducted a comparative analysis of three major AI chatbots: ChatGPT, Microsoft Copilot, and Gemini. Three clinical vignettes, one for each KDT type, were created to represent common KDT diet plans: the Classic Ketogenic Diet (Classic KD), the Medium-Chain Triglyceride (MCT) diet, and the Modified Atkins Diet (MAD). Two KDT dietitians evaluated the AI-generated diet plans for accuracy of macronutrient calculations, appropriateness of food choices, and overall safety and feasibility. RESULTS: All three chatbots failed to produce accurate KD plans. In the Classic KD, none of the chatbots generated menus that matched their calculated macronutrient targets, resulting in widely varying and inaccurate ketogenic ratios. For the MCT diet, the generated menus provided significantly less MCT than prescribed and failed to meet the caloric and macronutrient goals. For the MAD, the chatbots exceeded the specified net carbohydrate limits and provided inappropriate food suggestions. CONCLUSION: AI chatbots, in their current state, are not reliable or safe for generating KDT plans for children with epilepsy. The inaccuracies in macronutrient calculations and inappropriate food suggestions pose significant risks. This study underscores the critical need for human oversight by qualified professionals in the clinical application of AI for specialised dietary therapies.