On the Use of Personalized Models for Blood Glucose Concentration Prediction

Patients with type-l diabetes need to constantly monitor blood glucose concentration (BGC) level to stay in a healthy range. Consumer devices for BGC monitoring can be integrated with machine and deep learning techniques so that glucose level can be forecast and promptly provided to the patient. Recent advancements in the field suggest the use of a cus-tomization step based on each subject for blood concentration prediction. However, there is no comparison with other cus-tomization strategies and more importantly, there is no quan-titative analysis on the benefits of such a customization. In this paper: (1) we evaluate the impact of several pre-processing strategies on the performance; (2) we conduct a comparative analysis between 2 different customization methods and a general purpose strategy with no customization at all, and finally, (3) we propose a new personalization technique, called Threetask, that performs slightly better than other strategies on the majority of the patients, especially in the 60- and 90-minutes horizon. Experiments have been conducted on the OhioT1DM dataset which contains eight weeks of continuous monitoring of Blood Glucose Concentration from 12 subjects.