This thesis aims to develop individualized data augmentation strategies for type 1 diabetes based on personal eating patterns. It begins with an analysis of a real-world dataset to characterize meal distributions (breakfast, lunch, dinner, snacks) across pediatric age groups.

A MATLAB interface is developed to define hyperparameters for Gaussian Mixture Models (GMMs), which model each subject’s meal timing. The resulting GMMs are linked to specific meals, forming a personalized library.

Each test subject is matched to the most similar training subject based on their GMM. Meal generators are then built using the matched subject’s model to produce realistic, individualized meal patterns.

These synthetic meals are used to simulate glucose traces, from which glycemic metrics—such as time in, below, and above range—are extracted. The goal is to assess whether increasing synthetic data leads to convergence of glycemic outcomes, supporting more robust, personalized modeling for diabetes management.