Abstract

Diabetes is a chronic disease that currently has no cure. However, in the last decade, life- changing technology for people with diabetes has advanced primarily due to new sensors that continuously measure glucose. Individuals with diabetes manually use this data for diabetes management, but the artificial intelligence community has seen the increase in data as an opportunity to work towards automating diabetes control. The project developed deep-learning neural networks incorporating blood glucose measurements measured by a continuous glucose monitor (CGM) and physical activity data from a wearable health sensor (WHS) to predict future glucose values and compare how different configuration variables, such as input length or health features, impacted prediction accuracy. Model accuracies were compared using root mean squared error (RMSE). The data used to train and test the prediction models was from the OhioT1DM 2018 dataset that contains physiological signals collected from a WHS and blood glucose measurements from a CGM. The dataset was processed and organized into input and output dimensions using a custom created configuration file. The prediction model used was a deep learning Long Short-Term Neural Network (LSTM). The model was trained on all participants but tested on each participant individually by comparing experimental and measured blood glucose predictions at 0.5 and 1.0 hours.The findings suggested increased input length could improve predictions, but additional health features did not. All the health features considered, including insulin dosing, unexpectedly decreased prediction accuracy. A novel approach that compared predictions from single-value and series RMSE calculations showed that the series output approach provided additional context about how the models fit the data. Future research should address how results could be compared across literature studies and focus on event-based feature extraction from WHS.

Library of Congress Subject Headings

Blood glucose monitoring--Technological innovations; Blood glucose--Forecasting--Automation; Deep learning (Machine learning)

Publication Date

4-21-2023

Document Type

Thesis

Student Type

Graduate

Degree Name

Mechanical Engineering (MS)

Department, Program, or Center

Mechanical Engineering (KGCOE)

Advisor

Kathleen Lamkin-Kennard

Advisor/Committee Member

Jason Kolodziej

Advisor/Committee Member

Jamison Heard

Campus

RIT – Main Campus

Plan Codes

MECE-MS

Download

Share

COinS