M.S. AAI Capstone Chronicles 2024

Commercial advancements have followed similar trends. Companies like Omnipod and Tandem have integrated CGMs with insulin pumps to provide real-time machine learning-based alerts for glycemic events (Ciprich, 2024). Medtronic previously used the Sugar.IQ system to predict glucose levels in 60 minutes, but discontinued this in 2022 in favor of a model that adjusts basal insulin rates every five minutes (Medtronic, 2024). Mobile applications such as mySugr and Carb Manager assist in tracking glucose levels and dietary intake but lack predictive automation. Meanwhile, food recognition systems similar to Google’s work on the Nutrition 5k dataset highlight the potential of deep learning for dietary analysis, laying the groundwork for innovative AI-powered food tracking applications (Thames et al., 2021). Despite advancements, existing solutions often lack integration, requiring manual inputs or focusing on specific aspects of diabetes management. This project aims to address these gaps by combining predictive glucose modeling, long-term HbA1c estimation, and automated meal analysis in a single system. To address these integration challenges, the project employs a carefully selected combination of advanced machine learning models. At the core of our glucose prediction system are Long Short-Term Memory (LSTM) networks, specialized RNNs designed to excel at capturing temporal dependencies in sequential data. Their gating mechanisms effectively address common challenges like vanishing gradients, making them particularly suitable for predicting both blood glucose trends and HbA1c levels from CGM data.

For meal-related data, two complementary vision-based approaches are implemented. Convolutional Neural Networks (CNNs), specifically MobileNetV3Large and YOLOv8

7

241