M.S. AAI Capstone Chronicles 2024

Managing diabetes has been a focus of both academic research and commercial innovation. Numerous efforts have explored predictive models and automated systems for blood glucose management, meal analysis, and long-term glycemic control. Several studies have employed LSTM networks and hybrid architectures to predict blood glucose trends using CGM data and insulin dosing history. These models address challenges such as data irregularities and variability in glucose responses (Zhang et al., 2024). Fine-tuning insulin dosing based on carbohydrate intake has been recognized as an important factor in glycemic control. Current guidelines recommend that prandial insulin dosing be tailored to the carbohydrate content of meals (Tascini, 2018). By incorporating food image-based carbohydrate content prediction, this project seeks to enhance insulin dosing precision and improve diabetes management. Some researchers have also explored CNNs and Vision Transformers (ViTs) for food image analysis, achieving high accuracy in predicting macronutrient content and glycemic load (Hui, 2024). Ando et al, (2019) created DepthCalorieCam, a mobile application to estimate caloric content of various foods with limited success. Lastly, Google’s work on Nutrition 5k and studies integrating CGM data with demographic information (Nguyen, 2021) further demonstrate the feasibility of comprehensive machine learning solutions for diabetes care. Transformer-based models have also emerged as powerful tools for predicting Hemoglobin A1C (HbA1c), offering interpretability and scalability in analyzing long-term glucose trends (Acuna et al., 2023). These methodologies are particularly effective in real-world applications, enabling automated meal assessments and long-term glucose monitoring. Zhang et al. (2024) employed multi-stage LSTM networks for real-time glucose trend prediction, significantly improving accuracy over traditional statistical models.

6

240