Bayesian optimization assisted meal bolus decision based on Gaussian processes learning and risk-sensitive control

Effective postprandial glucose control is important to management for subjects with diabetes mellitus. In this work, a data-driven meal bolus decision method proposed without the need of subject-specific parameters. The dynamics learnt using Gaussian process regression. Considering asymmetric risks hyper- and hypoglycemia uncertainties in predicted trajectories, an risk-sensitive cost function designed. Bayesian optimization utilized solve problem, since gradient unavailable. approach evaluated 10-adult cohort FDA-accepted UVA/Padova T1DM simulator compared standard insulin calculator. For case announced meals, achieves satisfactory similar performance terms mean percentage time [70, 180] mg/dL increasing risk hypoglycemia. Similar results are observed information (assuming that patient follows consistent diet) basal rate mismatches. addition, comparison run-to-run based scenario potentially incorrect carbohydrate counts also performed, show more robust counting disturbances. At last, advisory-mode analysis performed on clinical data, which indicates can determine safe reasonable boluses real settings. verify effectiveness robustness indicate feasibility achieving improved regulation through optimal method.