Carbohydrate counting in traditional Turkish fast foods for individuals with type 1 diabetes: Can artificial intelligence models replace dietitians?

Abstract

Objectives Carbohydrate counting is a recommended approach for achieving glycemic control in individuals with type 1 diabetes (T1D). This study aimed to compare the accuracy of carbohydrate content estimations for traditional Turkish fast foods made by artificial intelligence (AI) models and dietitian. Methods Children and adolescents with T1D were pretested to identify the 12 most preferred Turkish fast-food items. Standardized recipes were developed for these meals, and the meals were photographed under standardized angular and lighting conditions. The photos were then uploaded to AI applications (ChatGPT-4.0, DeepSeek, Gemini, and CarbManager) and each model was prompted to estimate the carbohydrate content of the respective food items. Dietitians were asked to estimate the carbohydrate content based on these photographs. Results Of the dietitians in the study ( n = 40), 50% had postgraduate education, and 17.5% of those providing carbohydrate counting education ( n = 20, 50.0%) had been doing so for more than 7 y. No significant difference was found between the carbohydrate estimates of dietitians who provided and those who did not provide carbohydrate counting training ( P > 0.05). The intraclass correlation coefficient (ICC) between the AI models was 0.3554 (95% confidence interval [CI]: 0.0974–0.6801), indicating low reliability. The highest agreement with the estimates of dietitians who provided carbohydrate counting training (ICC = 0.417, 95% CI: 0.247–0.685) and those who did not (ICC = 0.307, 95% CI: 0.163–0.578) was observed with ChatGPT. Conclusions AI models can assist individuals with diabetes and healthcare professionals in estimating the carbohydrate content of foods, and consequently, can make a significant contribution to diabetes self-management.