Glucose Prediction Using Population-based Models and Genetic Data in Type 1 Diabetes Patients

Type 1 Diabetes (T1D) is a chronic autoimmune disorder causing dys-glycaemia. Technological advancements, such as continuous glucose sensors and closed-loop insulin delivery systems, are potential solutions to more efficient disease management. Hyperglycaemia contributes to the development of long-term micro-or macro-vascular complications; on the other hand, hypoglycaemia remains an acute and potentially life-threatening challenge impacting the quality of life for individuals with T1D. This study is a prodromal analysis of how specific genetic variants related to hypoglycaemia can be introduced to short-term predictive modelling methods of subcutaneous glucose concentration prediction in people with T1D, aiming at improving prediction errors, especially, in hypoglycaemic regions. We employed a K-mode-based clustering of study participants based on their genetic data and we developed separate population-based univariate long short-term memory (LSTM) models per cluster. Cluster-wise population models, developed and tested on a dataset of 47 T1D patients, exhibited a MAPE of 11.6%, 12.8% and 12.0% on the 3 identified clusters, respectively, for a prediction horizon of 30 min. Cluster-wise predictions outperformed global population models for the case of Group 0 which possess an increased percentage of very low (<54 mg/dL) and very high (>250 mg/dL) glucose values. This finding shows that incorporating genetic data into the pipeline of population-based glucose predictive modelling in T1D can improve the prediction error.