Comparative accuracy of automated carbohydrate estimation using NutriDish and ChatGPT-5 against dietitian-derived reference values
TL;DR
NutriDish demonstrated good agreement with dietitian-derived carbohydrate estimates, highlighting its potential to aid carbohydrate counting in individuals treated with insulin.
Abstract
Background Accurate carbohydrate estimation is critical for insulin dosing in individuals with diabetes on multiple daily injections. Conventional methods such as visual estimation and food exchange systems are inconsistent, particularly in multicultural Asian settings where portion sizes vary and nutritional labeling is limited. NutriDish is an AI-powered system designed to automates carbohydrate estimation from food images. Objective This study compared the accuracy of carbohydrate estimates generated by NutriDish and ChatGPT-5 against reference values established by a dietitian using weighed food portions. Methods Between July and August 2025, food items were collected from hawker centers, food courts, restaurants, and homes, supermarkets, and convenience stores across Singapore to capture the region's diverse culinary landscape. Items were classified as simple foods or composite meals. NutriDish and ChatGPT-5 independently estimated carbohydrate content, while a dietitian established reference values using weighing portions and a validated nutrient database. Accuracy was assessed using Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), proportion of estimates within ± 10 g of the reference value, and Bland-Altman analysis. Results Across 150 food items, NutriDish achieved significantly lower MAE [5.9 ± 5.3 g (median: 4.1 g, IQR: 1.0–9.0) vs. 8.7 ± 7.4 g (median: 6.0 g, IQR: 3.0–13.4); p < 0.001] and MAPE [22.1 ± 35.7% (median: 13.0%, IQR: 4.9–25.0) vs. 32.9 ± 50.8% (median: 22.1%, IQR: 9.1–37.1); p < 0.001] than ChatGPT-5. NutriDish estimates fell within ± 10 g of the reference values in 80.7% of cases vs. 69.3% for ChatGPT-5 (p = 0.016). Simple food items were estimated more accurately than composite meals, with MAEs of 4.1 ± 4.3 g and 8.6 ± 5.6 g, respectively (p < 0.001). Regarding reference object types, nutrition labels yielded the highest accuracy (MAE: 2.7 ± 4.2 g; MAPE: 9.0% ± 14.5%), whereas coins were associated with the poorest performance (MAE: 9.3 ± 7.3 g; MAPE: 48.1% ± 74.6%). Conclusion NutriDish demonstrated good agreement with dietitian-derived carbohydrate estimates, highlighting its potential to aid carbohydrate counting in individuals treated with insulin. Future studies should evaluate real-world clinical usability, integration into diabetes self-management ecosystems and impact on long-term glycemic control.