M.S. AAI Capstone Chronicles 2024

architectures, extract hierarchical spatial features from food images with high accuracy, excelling at food identification and carbohydrate estimation. Additionally, Vision Transformers (ViTs) complement these models by dividing images into patches and applying self-attention mechanisms, enabling both global and local features analysis for glycemic load and macronutrient content prediction. LSTMs are particularly effective for modeling short-term fluctuations and long-term glucose trends, aligning naturally with the temporal nature of CGM data. Meanwhile, CNNs and ViTs enhance meal analysis by providing a detailed breakdown of food composition. Together, these models form a robust system for improving blood glucose prediction, optimizing insulin dosing, and automating dietary assessment. Existing research and commercial solutions validate the methodologies chosen for this study and provide benchmarks for model performance and system design. This project builds upon these foundations to develop an integrated, AI-driven diabetes management system that bridges the gap between glucose prediction, HbA1c estimation, and automated meal analysis. Experimental Methods This project employs three core machine learning models, each tailored to address specific tasks in diabetes management: The LSTM Model for Blood Glucose Prediction features two sequential LSTM layers, each with 128 hidden units, followed by a dense layer with 64 nodes with ReLU activation. A final dense layer outputs the predicted blood glucose value as a continuous variable. Dropout layers placed after each LSTM layer prevent overfitting. The input features include normalized blood glucose

8

242