A Multivariate Bayesian Optimization Framework for Long-Term Controller Adaptation in Artificial Pancreas

In this work, we consider the problem of longterm parameter adaptation in artificial pancreas (AP). A parameter adaptation layer that operates on a larger timescale is firstly introduced, on top of the real-time closed-loop glucose control algorithms. A multivariate Bayesian optimization (BO) assisted parameter adaptation framework is then proposed, which features a dynamic parameter selection module that adaptively selects the parameter to be optimized and a BObased optimization module that adjusts the parameter through optimizing an unknown cost function. The proposed parameter adaptation method is evaluated on the 10-patient cohort of the US Food and Drug Administration accepted Universities of Virginia/Padova simulator through two extreme in silico scenarios. In the first scenario, we show that the proposed method can efficiently reduce average glucose from 173.1 mg/dL to 138.0 mg/dL (p < 0:001) and improve percent time in the euglycemic range [70, 180] mg/dL from 63.9% to 93.2% (p < 0:001) without adding any additional risk of hypoglycemia. In the second scenario, the proposed algorithm is able to alleviate hypoglycemia in terms of percent time below 70 mg/dL, from 12.5% to 0.2% (p < 0:001), while improving percent time in [70, 180] mg/dL from 79.4% to 91.4% (p < 0:001). The obtained results indicate feasibility and efficiency of adopting BO-based algorithms in long-term AP adaptation.