Model Personalization in an Advanced Automated Insulin Delivery System: An In-Silico Exploration

It is well-known that tailored solutions are necessary to achieve tight glycemic control in diabetes care. In this work, we analyze the performance impact of model personalization on our clinically validated, fully Automated Insulin Delivery (AID) system, the so-called RocketAP. In this framework, personalization is focused on the glucose-insulin model used by its Kalman filter and Model Predictive Control (MPC) algorithm. To this end, identifiability analysis is conducted to obtain a suitable set of model parameters that are then estimated on a daily basis. Performance of RocketAP using a population-based or a personalized model is evaluated considering the 100 adult cohort of the Food and Drug Administration (FDA)-accepted UVA/Padova Type 1 Diabetes (T1D) Simulator. Comparing with the population-based model, results with the proposed personalization approach led to an increase in the percentage of time spent between 70–140 mg/dL (40.9% to 50.5%: +9.6%) and 70–180 mg/dL (70.5% to 75.0%: +4.5%), with an associated slight increase in the percentage of time spent below 70 mg/dL (0.6% to 1.6%: +1.0%). We conclude that model personalization has potential to improve glucose control overall in our fully AID system, RocketAP. However, further studies are needed to confirm the feasibility of this approach.