Black-Box Model Identification of Patients with Type 1 Diabetes for Short-Term Glucose Prediction
Résumé
Type 1 Diabetes Mellitus (T1DM) is a chronic autoimmune disease characterized by the destruction of pancreatic β cells, i.e. the cells responsible for producing the endogenous insulin needed for maintaining Blood Glucose (BG) within the range of 70–180 mg/dL. As a result, patients with T1DM require lifelong exogenous insulin therapy. Nowadays, the use of Automated Insulin Delivery (AID) systems is well established as the most promising therapeutic approach for BG regulation. Due to meal intake, physical activities, or physiological variations (e.g. stress), T1DM patients are, however, still subject to the risk of hypoglycaemia (BG <70 mg/dL) and hyperglycaemia (BG > 180 mg/dL) [1]. In this context, the introduction of BG prediction into control policy has received increasing attention [2], see e.g. the use of BG prediction models coupled with Model Predictive Control (MPC) setups in [3] or the ARXmodels in [4]. Such predictive capabilities additionally enable the anticipation of glycaemic events, allowing corrective action before glucose levels breach safety thresholds [5]. In this work, three Black-Box models for multi-step glucose prediction are identified using real patient data extracted from a cohort of 29 adult patients with T1DM [6]. Linear ARX, ARMAX (with moving-average noise) and Box–Jenkins (BJ) (with independent process and noise dynamics) models are investigated. Prediction performance is evaluated in terms of Root Mean Square Error (RMSE) and FIT for horizons of 5, 15, 30, and 60 minutes. All models are identified on a patient-specific basis to account for inter-patient variability.