Research Article
An Exploratory Study of the Value of Vital Signs on the Short-Term Prediction of Subcutaneous Glucose Concentration in Type 1 Diabetes – The GlucoseML Study
@INPROCEEDINGS{10.1007/978-3-031-34586-9_30, author={Daphne N. Katsarou and Eleni I. Georga and Maria Christou and Stelios Tigas and Costas Papaloukas and Dimitrios I. Fotiadis}, title={An Exploratory Study of the Value of Vital Signs on the Short-Term Prediction of Subcutaneous Glucose Concentration in Type 1 Diabetes -- The GlucoseML Study}, proceedings={Pervasive Computing Technologies for Healthcare. 16th EAI International Conference, PervasiveHealth 2022, Thessaloniki, Greece, December 12-14, 2022, Proceedings}, proceedings_a={PERVASIVEHEALTH}, year={2023}, month={6}, keywords={Type 1 Diabetes Glucose Prediction Machine Learning Continuous Glucose Monitoring Vital Signs}, doi={10.1007/978-3-031-34586-9_30} }- Daphne N. Katsarou
Eleni I. Georga
Maria Christou
Stelios Tigas
Costas Papaloukas
Dimitrios I. Fotiadis
Year: 2023
An Exploratory Study of the Value of Vital Signs on the Short-Term Prediction of Subcutaneous Glucose Concentration in Type 1 Diabetes – The GlucoseML Study
PERVASIVEHEALTH
Springer
DOI: 10.1007/978-3-031-34586-9_30
Abstract
The daily self-management of type 1 diabetes (T1D) has benefitted from the advancements in real-time continuous glucose monitoring and hybrid closed-loop insulin delivery. These technologies comprise, in parallel, significant sources of data providing insight into daily glucose control and insulin treatment. The concurrent real-time continuous monitoring of vital signs, 24/7, complements the exploitable information for one individual. In this study, we investigate whether respiratory, hemodynamic, and body temperature vital signs correlate linearly with the subcutaneous glucose concentration in T1D, and improve its short-term, up to 60-min ahead, prediction as compared to a univariate model. To verify our research hypothesis, (i) we approximate the prediction of glucose concentration via a long short-term memory (LSTM) function of the recent 30-min history of glucose and those vital signs with a Pearson’sr> 0.5, and (ii) contrast its performance with that of the univariate model. LSTM has been trained and tested individually, using a dataset with 22 T1D people monitored for 2 or 4 weeks. Our analysis identified that: (i) subcutaneous glucose concentration is linearly correlated principally with heart rate and systolic blood pressure, and (ii) the value of the vital signs lies in the improvement of the predictions in hypoglycaemia as the prediction horizon (PH) increases, where we observed a substantial reduction of erroneous predictions from 19% to 7% for a PH of 60 min.
