Degree
Doctor of Philosophy
Program
Computer Science
Supervisor
Dr. Charles Ling
Abstract
Self-Monitoring of Blood Glucose (SMBG) for Type-2 Diabetes (T2D) remains highly challenging for both patients and doctors due to the complexities of diabetic lifestyle data logging and insufficient short-term and personalized recommendations/advice. The recent mobile diabetes management systems have been proved clinically effective to facilitate self-management. However, most such systems have poor usability and are limited in data analytic functionalities. These two challenges are connected and affected by each other. The ease of data recording brings better data for applicable data analytic algorithms. On the other hand, the irrelevant or inaccurate data input will certainly commit errors and noises. The output of data analysis, as potentially valuable patterns or knowledge, could be the incentives for users to contribute more data.
We believe that the incorporation of machine learning technologies in mobile diabetes management could tackle these challenge simultaneously. In this thesis, we propose, build, and evaluate an intelligent mobile diabetes management system, called GlucoGuide for T2D patients. GlucoGuide conveniently aggregates varieties of lifestyle data collected via mobile devices, analyzes the data with machine learning models, and outputs recommendations.
The most complicated part of SMBG is diet management. GlucoGuide aims to address this crucial issue using classification models and camera-based automatic data logging. The proposed model classifies each food item into three recommendation classes using its nutrient and textual features. Empirical studies show that the food classification task is effective.
A lifestyle-data-driven recommendations framework in GlucoGuide can output short-term and personalized recommendations of lifestyle changes to help patients stabilize their blood glucose level. To evaluate performance and clinical effectiveness of this framework, we conduct a three-month clinical trial on human subjects, in collaboration with Dr. Petrella (MD). Due to the high cost and complexity of trials on humans, a small but representative subject group is involved. Two standard laboratory blood tests for diabetes are used before and after the trial. The results are quite remarkable. Generally speaking, GlucoGuide amounted to turning an early diabetic patient to be pre-diabetic, and pre-diabetic to non-diabetic, in only 3-months, depending on their before-trial diabetic conditions. cThis clinical dataset has also been expanded and enhanced to generate scientifically controlled artificial datasets. Such datasets can be used for varieties of machine learning empirical studies, as our on-going and future research works.
GlucoGuide now is a university spin-off, allowing us to collect a large scale of practical diabetic lifestyle data and make potential impact on diabetes treatment and management.
Recommended Citation
Luo, Yan, "Machine Learning of Lifestyle Data for Diabetes" (2016). Electronic Thesis and Dissertation Repository. 3650.
https://ir.lib.uwo.ca/etd/3650