Related Experiment Video
Updated: May 2, 2026
'Boden Food Plate': Novel Interactive Web-based Method for the Assessment of Dietary Intake
Published on: September 18, 2018
Comparative accuracy of smartphone apps and a generative AI tool for carbohydrate counting: An independent bicentric
Michael Joubert1, Bleuenn Dreves1, Theo Arnould1
1Diabetes Care Unit, Caen University Hospital, UNICAEN, Caen, France.
Aims:
Accurate carbohydrate counting is essential for optimal prandial insulin dosing in individuals with type 1 diabetes (T1D), including hybrid closed-loop (HCL) users. Although several applications assist with real-time carbohydrate estimation, few independent studies have compared their accuracy head-to-head. This study aimed to evaluate, in a standardized hospital-based environment, the real-world carbohydrate estimation accuracy of four dedicated smartphone applications-DiabHealth®, GluciCheck®, EkiYou®, Gluroo®-and, for the first time, the multimodal generative AI ChatGPT-5.
Materials And Methods:
In this prospective bicentric study, trained medical students acting as mock patients estimated the carbohydrate content of 246 hospital meals using the five tools. Reference values were generated by expert dietitians based on weighed food items and standardized photographs. The primary endpoint was absolute error in carbohydrate estimation. Secondary outcomes included the proportions of meals estimated within ±10 and ±20 g. Agreement analyses used Spearman correlations and Bland-Altman methods.
Results:
Mean absolute errors were 13.0 ± 9.9 g for GluciCheck, 14.2 ± 12.5 g for EkiYou, 13.9 ± 9.9 g for DiabHealth, 20.6 ± 14.6 g for Gluroo, and 18.0 ± 16.0 g for ChatGPT-5. Using GluciCheck as reference, only Gluroo showed significantly higher error (p < 0.05). Spearman correlations confirmed strongest associations for GluciCheck (ρ = 0.78) and EkiYou (ρ = 0.71), with moderate correlations for DiabHealth, Gluroo, and ChatGPT-5 (ρ ≈ 0.50). Bland-Altman analyses showed the narrowest limits of agreement for GluciCheck and EkiYou, and the widest for Gluroo. DiabHealth and ChatGPT-5 showed minimal bias and no proportional error. The proportion of meals estimated within ±20 g was 77% for GluciCheck, 67% for EkiYou, 80% for DiabHealth, 50% for Gluroo, and 73% for ChatGPT-5.
Conclusions:
In this controlled hospital-based evaluation, GluciCheck, DiabHealth, and EkiYou provided the most accurate and consistent carbohydrate estimations. ChatGPT-5, despite not being designed for nutritional analysis, achieved intermediate accuracy comparable to several dedicated tools, whereas Gluroo showed the greatest variability. These findings highlight the potential of both structured manual systems and emerging AI-based technologies to support carbohydrate counting in T1D management.
