Improving Blood Glucose Prediction for People with T1DM During Physical Activity Using Machine Learning on Participant Collected Data

Abstract

For people with Type 1 Diabetes Mellitus (T1DM), engaging in physical activities (PA) presents unique challenges. The aim of this thesis was to improve the prediction of blood glucose (BG) levels for individuals with T1DM during and after PA. The study began with a literature review to guide the research direction and understand existing prediction models. Then particular emphasis was placed on analyzing papers that provided open-source code, allowing validation of these models using the OhioT1DM dataset and data collected from participants. The GluPredKit platform, an open-source blood glucose prediction framework, was used to streamline the process of data handling, training, and evaluating BG prediction models in Python. The study progressed by training and evaluating various machine learning (ML) models with data from two participants with T1DM. Finally, Physiological Hybrid models and various ensemble models were implemented to observe performance improvement during physical activities.

The Physiological Hybrid model did not improve the predictions during PA compared to the conventional ML models. Although ensemble modeling provided a slight improvement in prediction performance, no ensemble consistently outperformed others, indicating a need for further refinement. Additionally, traditional metrics like Root Mean Squared Error (RMSE) were found to be insufficient in accurately assessing model performance during PA. This prompted the introduction of an additional evaluation method, trajectory plots.

Despite these advancements, this study has several limitations, including the small sample size and heavy reliance on data from smartwatches. As a result, future research should focus on recruiting more participants, refining metrics to better assess ML model performance during PA, and exploring innovative modeling approaches to achieve improved outcomes.