PhD thesis defense to be held on December 20, 2023, at 15:00 (Teleteaching Room, NTUA Central Library)

Picture Credit: Stavros Pitoglou

Thesis title: Combination of Dynamic System Analysis and Machine Learning techniques to explore new approaches in the development and verification of personalized models of short-term Glycemic Dynamics prediction, and development of a standard software platform aimed at self-management of Diabetes Mellitus utilizing Cloud and Mobile Computing Technologies

Abstract: Chronic diseases such as Diabetes Mellitus (DM) are a constantly growing problem for modern societies, regardless of their stage of development and socio-economic situation. DM in particular, unlike other diseases, displays increased dynamics in developing and developed social formations, which makes its prevention, treatment and management urgent and imperative.
From the point of view of biomedical technology, one of the key challenges is the personalized short-term prediction of glucose concentrations in human blood, in order to prevent glycemic deregulation and provide a scientifically based basis for timely compensatory measures. Maintaining euglycemia, i.e. glucose concentrations around normal levels, is a critical way of managing the disease, especially when one takes into account the significant influence exerted on the involved physiological mechanisms by everyday events such as meal ingestion, physical activity, etc.
Regarding the machine learning methods that can be used for short-term prediction purposes, this dissertation provides a critical review of the existing literature and proposes a specific methodological approach that is based on: a) the assumption that the time series of glucose concentration is the superposition of the result of all the regulatory and counter-regulatory mechanisms that influence glycaemic factors., b) the personalized approach resulting from the dynamic selection and training of models whose final characteristics differ from person to person, but also for the same person in different time periods. Also, a new metric is proposed for estimating the predictive performance of models, which in a single mathematical formulation contains combined elements of numerical and clinical evaluation.
As part of this approach, a personalized modeling process based on machine learning techniques was created, which yielded a result clinically comparable to the corresponding optimal research efforts reported in the literature, while having great margins for improvement. This effort aspires to provide a more solid foundation for the methodological approach to the problem of specific predictive modeling and for its clinical application in real-world conditions.
Finally, a prototype hardware and software combination system was developed, which was based on cloud computing architectures and the use of portable smart devices, in order to implement the dynamic modeling and short-term prediction process proposed and provide the basis for personal information solutions as well as automated management functionalities such as, for example, artificial pancreas devices.

Supervisor: Professor Emeritus Dimitrios - Dionysios Koutsouris

PhD Student: Stavros Pitoglou