MPC Model Individualization in Free-Living Conditions: A Proof-of-Concept Case Study

In the last years, Model Predictive Control (MPC) proved to be one of the most promising approaches for an Artifcial Pancreas (AP), a device for closed-loop blood glucose control in subjects afected by Type 1 Diabetes (T1D). MPC performance is highly infuenced by the quality of the model used for prediction. Moreover, the inter-patient variability characterising subjects with T1D increases the need of patient-tailored models. Recently, promising results have been obtained in silico using the UVA/Padova simulator in Soru et al. (2012) and Messori et al. (2016) where diferent individualization techniques have been studied and compared to the “average” model of the UVA/Padova adult population showing signifcant improvements in term of prediction ability. The aim of this paper is to verify the applicability of the technique described in Soru et al. (2012) and extend it to be used on free-living data collected without ad hoc clinical protocols. Data were collected during a 1 month trial in free-living conditions (Renard et al. (2016)). In this proof-of-concept case study, individualized models obtained with diferent identifcation parameters are compared with the “average” model that was used to synthetize the MPC controller used during that trial. The individualized models show superior prediction performance and prove robustness to non-optimal algorithm initialization in a selected test-case.