Reliable artificial pancreas (AP) systems require accurate models of the glucose-insulin dynamics in people with type 1 diabetes mellitus (T1DM). The realization of a fully automated AP system, however, is not possible with manual user entries for meals and exercise. To promote the development of automated AP systems without any user inputs, an adaptive and personalized data-driven model is developed that characterizes the evolving glycemic dynamics of individuals while explicitly considering the estimated effects of unknown disturbances such as meals and exercise. To achieve this adaptive and personalized modeling approach that best describes the glucose dynamics at any given instant without manual user input, we first quantify the effects of the unknown meals through adaptive estimates of time-varying parameters in dynamic physiological model. The evaluated meal effects are incorporated with plasma insulin concentration estimates and readily measured physiological data from noninvasive sensors, such as heart rate and energy expenditure, to determine the overall state of the individuals. The proposed modeling approach integrates first-principles physiological models and data-driven empirical techniques to determine an accurate and comprehensive perspective of the individual at any given time. Simulation case studies involving in silico subjects from the U.S. Food and Drug Administration approved University of Virginia/Padova metabolic simulator demonstrate the improvement in prediction ability of the proposed approach. The root-mean-square error and mean absolute error for 30-, 45-, and 60-min-ahead predictions are 15.61 and 8.98 mg/dL, 21.82 and 13.49 mg/dL, and 25.96 and 17.00 mg/dL, respectively. The accurate predictions of future glucose concentration measurements without requiring manual user inputs will enable fully-automated AP systems that perform efficiently by maintaining euglycemia and mitigating glucose excursions.
I. Hajizadeh: None. M. Rashid: None. A. Cinar: None.