Image-based Estimation of Dietary Sugar Intake in Taiwanese Student Meals for Metabolic and Kidney Health Surveillance

Excessive dietary sugar intake is a growing public health concern in East Asia and is linked to metabolic syndrome, obesity, insulin resistance, and long-term kidney risk. In Taiwan, over 95% of junior-high students consume at least one sugar-sweetened beverage daily, and nearly one-quarter report weekly energy-dense, ultra-processed foods consumption. Given the strong association between high sugar consumption and early-life metabolic dysregulation, there is a need for tools to quantify sugar exposure in real-world dietary settings.

We developed a prototype system based on a pretrained convolutional neural network (CNN) and fine-tuned it using the Food-101 dataset for food and beverage image recognition. To enable dietary sugar estimation, we manually calculated the average sugar content for each of the 101 food categories in the dataset. The system processes a photograph of a student’s meal or drink, identifies the items using the fine-tuned model, and estimates their average sugar intake when corresponding data are available. Given growing concerns about excessive sugar consumption among students, this prototype targets real-world dietary habits. For instance, a recent survey reported that popular sugar-sweetened beverages in Taiwan exceeded WHO sugar-intake guidelines in over 40% of cases. To evaluate our approach under realistic conditions, we collected photographs of student meals—including packaged lunch boxes and drinks—from school canteens and convenience stores for preliminary validation.

By applying our method to everyday meals, the system demonstrated a reasonable recognition rate for food types included in the Food-101 dataset. This capability allows it to estimate the total amount of dietary sugar in a meal with corresponding reliability, based on the identified food categories. However, the model’s accuracy depends heavily on how closely the food images align with the categories in the training dataset. The system’s performance declined when analyzing food items not represented in the Food-101 dataset or meals composed of mixed dishes, where visual boundaries between components were unclear. This limitation is particularly evident for beverages with ambiguous portion definitions, a common issue in Taiwan’s drink market. Nevertheless, given the high baseline consumption of sugary beverages among students, our approach shows promise for identifying high-sugar items and supporting early dietary interventions to help prevent long-term health issues in Taiwanese students.

This study demonstrates the feasibility of an image-based approach for estimating sugar intake from student meals—an especially relevant application in Taiwan, where sugar-sweetened beverages and processed lunches are common in youth diets. Future work will focus on improving classification accuracy for mixed meal types, expanding the training dataset to better represent local Taiwanese lunch-box items, and validating the system in a larger student cohort. Ultimately, such a tool could aid dietary monitoring and support health-promotion initiatives in schools.