Department of Computer Science: MSc Thesis Presentation
Exploring latent structure in Continuous Glucose Monitor data
Author: Viljami Raiski
Supervisor: Juho Rousu
Date: Friday 7 May 2021
Diabetes mellitus is a growing problem in both, developing and developed countries, affecting the lives of over 422 million people worldwide. The traditional way of diagnosing and detecting abnormal blood glucose behaviour has been via single-time point measurements, but nowadays arrival of continuous glucose monitoring (CGM) systems have enabled access to the blood glucose time series.
Blood glucose time series provided by CGM devices have been utilised in numerous use-cases such as forecasting blood glucose values and classifying diabetes via supervised learning methods, detecting abnormal patterns via unsupervised learning methods and designing blood glucose control algorithm by using reinforcement learning methods.
In this thesis, we explore the applicability of supervised learning methods for capturing the latent structure in the individual blood glucose dynamics. Furthermore, we experiment with the usefulness of the captured latent structure in two use cases, classification between subjects diagnosed with type 2 diabetes (T2D) and prediabetes, and high blood sugar prediction.
The thesis suggests that regression coefficients could capture relevant information about individuals blood glucose dynamics but further research and clinical validation is required due to small sample size, possible inaccuracies in the CGM technology and lack of comparison between current medically validated methods for assessing individuals blood glucose dynamics.