Background: Accurate carbohydrate counting is fundamental to glycaemic management in type 1 diabetes (T1D), but visual estimations are cognitively demanding and error prone. Image-based nutrition apps that use artificial intelligence for food recognition and volume estimation may reduce burden, yet their accuracy across diverse meal types and sizes is not well established.
Objective: To compare the performance of image-based nutrition tracking apps to estimate carbohydrate content of meals against photo-based visual estimates of adults with type 1 diabetes.
Methods: Forty-eight adults with T1D and four commercial nutrition apps (SnapCalorie, MyFitnessPal, LoseIt, and MacroFactor) analysed ten dishes served in three meal sizes (small, medium, large). Each dish was matched for macronutrient composition (as % of total kJ), with fiducial markers provided for scale. Participants completed the task remotely from paired aerial and side-view photographs using an online survey platform. In total, 2,039 meal assessments were compared with weighed reference values (apps, n = 491; T1D, n = 1,548). Level of agreement was evaluated using Bland–Altman analysis (bias and 95% limits of agreement, LoA) and mean absolute error (MAE). Clinical significance was defined as MAE > 20 g.[1] Between group comparisons were examined using Mann–Whitney tests for MAE and chi-square tests for proportions.
Results: Mean bias for small meals shifted from +10.5 g (95% LoA: -27.5 g, +48.6 g) for apps and +1.1 g (95% LoA: -37.8 g, +39.9 g) for adults with T1D to increasing underestimation as meal size increased [medium meals: -1.3 g (95% LoA: -43.2 g, +40.6 g) for apps and -3.7 g (95% LoA: -51.5 g, +44.1 g) for participants; large meals: -15.6 g (95% LoA: -58.5 g, +27.3 g) for apps and -10.2 g (95% LoA: -58.9 g, +38.5 g) for participants]. For all meal sizes, MAE was within the ±20 g clinical limit for both assessment methods, however, apps had a significantly larger MAE for small meals (17.7 g versus 13.9 g; P < 0.001). Overall, 36.8% of all estimates exceeded the ±20 g threshold, and the difference was not statistically significant between assessment methods (apps 39.3% vs humans 36.0%, χ² = 1.704, P = 0.19).
Conclusion: Nutrition-tracking apps achieved similar overall accuracy to photo-based carbohydrate estimations of adults with T1D. As meal size increased, estimates trended towards underestimation, and variability increased. Because about one third of individual errors exceeded the ±20g threshold, image-only estimates (either apps or visual estimates from photos) could expose adults with T1D to clinically significant errors in carbohydrate estimation.