Computer vision-based carbohydrate estimation for type 1 patients with diabetes using smartphones

Area of Science:

• Biomedical Engineering
• Medical Informatics
• Nutritional Science

Background:

• Type 1 diabetes (T1D) management requires accurate carbohydrate (CHO) counting for prandial insulin dosing.
• Inaccurate CHO estimation, even by 20 grams, significantly impacts postprandial glucose control.
• Manual CHO counting is challenging for T1D patients.

Purpose of the Study:

• To develop and evaluate the GoCARB system, a smartphone application for automated carbohydrate counting in non-prepackaged foods for T1D patients.
• To assess the accuracy of the GoCARB system in estimating carbohydrate content from food images.

Main Methods:

• The GoCARB system utilizes smartphone images of a meal with a reference card.
• Image analysis includes plate detection, food segmentation, recognition, and 3D shape reconstruction.
• Carbohydrate content is calculated using estimated food volumes and the USDA nutritional database.

Main Results:

• Laboratory evaluation used 24 multi-food dishes with multiple image pairs and system applications per dish.
• The system achieved a mean absolute percentage error of 10 ± 12% in CHO estimation.
• The mean absolute error for normal-sized dishes was 6 ± 8 grams of CHO.

Conclusions:

• The GoCARB prototype system demonstrated feasibility, with errors below the 20-gram target.
• Further research and development are necessary before clinical deployment to accommodate diverse eating habits.
• The system shows promise for improving self-management in type 1 diabetes.