Periodic predictor of glucose dynamics in type 1 diabetes patients

Glucose dynamics in type 1 diabetes are highly variable both across individuals and within the same individual throughout the day, due to factors such as meals, insulin sensitivity, and daily routines. This variability poses significant challenges for accurate prediction and control, limiting the effectiveness of single-model approaches. The goal is to develop control-oriented models to provide accurate predictions of glucose dynamics, to be used within control strategies, such as model predictive control. The proposed approach adopts a multiple-models structure that accounts for both individual differences, in terms of metabolic response, and daily fluctuations in patient dynamics. Models are identified in different daily periods using an impulse response method applied to data generated by the UVA/Padova simulator. The day is divided into three time segments corresponding to breakfast, lunch, and dinner, with a separate model trained for each period. These models are integrated through a soft-switching mechanism giving rise to a multiple model predictor. Two state estimation techniques, the periodic Kalman filter and the moving horizon estimator, are compared to enable multi-step glucose prediction. Results indicate that the multiple model predictor consistently outperforms single-model approaches across the entire virtual adult population. This modeling framework shows strong potential for integration into advanced insulin delivery systems based on predictive control.