M.S. AAI Capstone Chronicles 2024

c.​ Nutrition 5k Dataset (Glycemic, 2024; Deeb, 2017): Consists of 3,492 meal photographs with nutritional information, supporting glycemic load prediction and carbohydrate content estimation for insulin dosing in T1D patients.

The project aims to develop a multi-component system that will do the following: a.​ Predict Blood Glucose levels; utilizing LSTM/GRU models to forecast glucose levels and recommend insulin doses based on patient-specific data. b.​ Predict HbA1c levels; analyzes glucose trends to estimate long-term control. c.​ Meal Analysis System; predicts carbohydrate content and glycemic load from meal photographs using CNNs and Vision Transformers. Together, these components provide a holistic toolkit for improving glycemic control, increasing HbA1c target outcomes, and ultimately to reduce complications associated with diabetes as illustrated in figure 1.1. This project leverages three datasets to support comprehensive diabetes management: 1.​ DiaTrend Dataset: Developed by Dartmouth Health, this longitudinal dataset includes CGM and insulin pump data from 54 participants (Eyth, 2023). It provides insights into real-world diabetes management, capturing blood glucose (BG) levels (mg/dl), insulin doses, and carbohydrate intake. Additional features, such as insulin-to-carbohydrate ratios and HbA1c values, are essential for predicting long-term glucose control. 2.​ Tidepool Dataset: A proprietary dataset from the Big Data Project, containing three months of CGM and insulin pump data (Tidepool, 2024). It includes glucose values, Dataset Summary

3

237