Date of Award

Spring 6-1-2021

Document Type

Thesis (Undergraduate)

First Advisor

Temiloluwa O. Prioleau


Carbohydrate counting, which refers to estimating the carbohydrate content in meals, is critical for determining mealtime insulin doses and maintaining healthy blood glucose levels in persons with type 1 diabetes (T1D). However, carbohydrate counting errors (i.e., over-or under-estimation of carbohydrate intake) are very common amongst patients and are often a source of poor glycemic control. Fortunately, the prevalence of personal health data from wearable medical devices like continuous glucose monitors (CGMs) and insulin pumps provide unique opportunities for understanding and predicting health management outcomes. In this study, we use adverse glycemic events following meal intakes as a proxy for identifying carbohydrate counting errors, then use supervised machine learning models to predict these carbohydrate counting errors. Our dataset includes an average of 161-days of CGM and insulin pump data from 34 patients with T1D. Using a total of 13 features from both datasets, we observed the highest prediction accuracy of 70.5% with a multilayer perceptron (MLP) classifier compared to a baseline model that only yielded 61% accuracy. This work provides a framework for the development of more data-driven tools that leverage personal health data for decision-support to improve health outcomes for people with T1D.