Inverse Optimal Control Using A Neural Multi-Step Predictor For T1DM Treatment

Regarding Type 1 Diabetes Mellitus (T1DM) research, the most recent advances in computer technology and biological sciences have focused on reducing complications due to hyperglycemic and hypoglycemic events. These dangerous conditions can lead to severe micro- and macrovascular damages, diabetic coma and even death. In this work, a novel inverse optimal approach for glycemic control in T1DM is proposed. The complete scheme integrates a neuronal identifier and a neural multi-step predictor to obtain future glucose readings; in order to allow that an optimal controller anticipates against a potential risk scenario. Validation is carried out using the well-known UVa/PAdova Simulator and the training data is obtained from a Continuous Glucose Monitoring (CGM) sensor. Closed-loop control results are presented with the average value for the three diabetic populations available in the UVa[Padova simulator: adults, adolescents and children. A basal insulin infusion rate and a five meal protocol are proposed for each population. For this scenario, the Neural Multi-Step Predictor (NMSP) configuration, combined with the neural inverse optimal controller, is able to keep glucose within normal levels 70-115 mg/dL and below 250 mg/dL postprandial lapse. Finally, the comparative framework of the Control Variability Grid Analysis (CVGA) displays that most of the patients are in the safe zones with the NMSP approach.