M.S. AAI Capstone Chronicles 2024

basal and bolus insulin rates, patient demographics, and device metadata, making it critical for developing real-time glycemic event prediction and insulin dosing models. 3.​ Nutrition 5k Dataset: Provided by Google Research, this dataset comprises food images and associated nutritional labels for over 5,000 meals (Thames, 2021). To test the calorie distribution of the different dishes, we first categorized the total calories into three bins: [Low, Medium, High]=[0,200,500]. Figure 3.1 shows that the Nutrition5k Dataset has sufficient representation for the different calorie categories. Key data features include carbohydrate, protein, fat content, and total calories. This dataset enables carbohydrate content estimation, essential for meal-time insulin bolusing and glycemic load prediction from meal photographs, offering key insights into dietary impacts on glucose levels. While exploring our chosen datasets we discovered some data issues and had to implement some preprocessing methodologies to prepare our data for training. For example, some of the metadata fields in the DiaTrend dataset were empty, and device settings were excluded when 100% null or irrelevant to the predictive task. The Nutrition 5k dataset contained some metadata rows without corresponding images, and these were removed to ensure the integrity of the image-label pairs. The Nutrition 5k dataset also contained some labels of ingredients that were essentially invisible, such as salt, olive oil, vinegar, etc. of which the camera would not reliably detect. These items often dominated in the validation set, exacerbating class imbalance, which adversely affected model inference, therefore they were excluded from the training dataset. We also found irregularities and noise in the DiaTrend dataset. Large Together, these datasets address unique and complementary management domains, enabling the development of an integrated system for diabetes management.

4

238