Open Access
Freeman, Kenneth Allan
Graduate Program:
Mechanical Engineering
Master of Science
Document Type:
Master Thesis
Date of Defense:
July 14, 2011
Committee Members:
  • Dr Qian Wang, Thesis Advisor
  • Adaptive Dual Control
  • Type-1 Diabetes
  • Artificial Pancreas
  • Automatic Control
A linear discrete time-varying model utilizing recursive parameter estimation coupled with an adaptive dual controller is presented for real-time modeling and control of the compromised glucose regulatory system in patients with type-1 diabetes mellitus. The performance of the proposed adaptive dual control for insulin delivery is evaluated via simulations using various performance indices widely adopted in existing literature. Compared to the clinical data (which include clinical measurement of glucose level, clinically-administered insulin delivery, and carbohydrate intake) collected from several patients over 72 hours, the simulation results show that the adaptive dual controllers developed in this thesis provide lower occurrence of hyperglycemia and hypoglycemia. In simulations, a virtual patient in the form of a linear time-varying input-output model was built based on clinical data using Kalman-filter based recursive parameter estimation. This virtual patient is then used in the simulations to evaluate the proposed adaptive dual controller for insulin delivery. The clinical data include measurement of glucose level, estimated carbohydrate intake and insulin delivery rate for five patients over 72 hours in 5-minutes intervals. Due to temporary interruptions in the data, the longest stretch of uninterrupted data was used, of which 49 hours was the minimum. Each of the patients was diagnosed with type-1 diabetes and under continuous subcutaneous insulin infusion plus bolus treatment. The virtual patient model takes into account the absorption and transport time delays existed in the subcutaneous insulin injection and carbohydrate intake in its design of the finite impulse filters for the system inputs. In evaluation of the virtual patient model, model correlation coefficients which are generated between half-hour future blood glucose prediction and clinical data glucose measurement, ranged from 0.84 to 0.96, showing sufficient model accuracy. Based on Kalman-filter estimation of system parameters from the virtual patient (the algorithm itself does not depend on the model of the virtual patient, and hence can be applied to any simulation environments as well as real clinical environments), an adaptive dual controller is designed to determine the insulin injection based on the feedback information of the glucose measurement, together with the estimated carb intake. The adaptive dual control algorithm minimizes two cost functions in the calculation of the control input in an attempt to cautiously track the target value while simultaneously providing persistent excitation required for accurate parameter estimation. Simulations show that the adaptive dual controller developed in this thesis has better control of type-1 diabetes with statistical significance when compared with the clinical treatment used during data acquisition. The results show that the adaptive approach based on real-time model estimation coupled with dual control could be a potentially very promising tool for closing the loop in blood glucose control in those with type-1 diabetes. Compared to the conventional approaches based on compartmental models and non-adaptive control designs (either classical PID controllers or modern optimal control designs) in the current literature, the proposed empirical modeling based on online recursive parameter estimation, together with adaptive control design has several advantages. First, the modeling and control designs can be used for patients in a natural living condition, with meal intake and exercises, noting the lack of a good meal model in the existing compartmental-model based control designs. Second, the proposed adaptive control approach provides the ability to track the time-varying behavior in a diabetic patient and react fast. Finally, the proposed framework allows easy extension to taking into account various factors that could affect patient’s diabetic control such as exercises, stress levels, etc in the future study.