(585e) The Extended and Unscented Kalman Filtering Methods for Real-Time Plasma Insulin Concentration Estimation in an Artificial Pancreas Control System for Patients with Type 1 Diabetes

Authors: 
Hajizadeh, I., Illinois Institute of Technology
Cinar, A., Illinois Institute of Technology
Cengiz, E., Yale University School of Medicine
Turksoy, K., Illinois Institute of Technology
Background and objective:Artificial pancreas (AP) systems use continuous glucose measurements (CGM) to calculate optimum amount of insulin to be infused with an insulin pump [1, 2]. Real-time estimations for plasma insulin concentration (PIC) rather than generalized curve based insulin-on-board (IOB) estimates can be beneficial for improving the performance of AP control algorithms to calculate more realistic insulin infusion rates and prevent hypoglycemia that would be caused by overdosing of insulin [3]. Our objective is to fulfill a real-time estimation of PIC from CGM data by using two different mathematical models.

Methods:Two different glucose-insulin compartmental models, Hovorkaâ??s model and extended Bergmanâ??s minimal model, which were developed to describe glucose-insulin dynamic in different parts of the human body, have been incorporated into a continuous-discrete extended Kalman filter (CD-EKF) and an unscented Kalman filter (UKF), respectively, to provide an estimate of the plasma insulin concentration [4,5]. Furthermore, because of variability in system dynamics, uncertain parameters have been considered as new states in Hovorkaâ??s model and Bergmanâ??s minimal model to be estimated by CD-EKF and UKF, correspondingly. Thirteen datasets from nine different subjects with type 1 diabetes are used. Two euglycemic clamps were performed on separate mornings with (seven datasets) and without (six datasets) an insulin infusion site warming device. Subjects had one CGM sensor and blood was collected to measure plasma insulin levels for up to 5 hr. On both days, a bolus of 0.2 U/kg aspart insulin was infused at the beginning of each experiment and the basal infusion was suspended, and glucose levels were maintained between 90 and 100 mg/dl by infusion of variable rate of dextrose [6]. In addition, partial least squares models based on each patientâ??s demographic information such as body weight, height, BMI and total daily insulin dose are developed for the initial guess of the time-varying unknown model parameters used in the nonlinear CD-EKF and UKF estimators.

Results: According to the CD-EKF estimator, the average root mean square errors are found to be 12.36 and 11.56 mU/L for the clamps with and without IP, respectively. The average mean absolute relative error with respect to the measured samples are 0.14 for the two clamps. The average Pearson's R coefficient are 0.95 and 0.91 for the clamps with and without IP, correspondingly. For UKF estimator, estimated plasma insulin concentrations are found to be linearly correlated with the measured insulin samples (R= 0.94 and 0.89 for the clamps with and without IP, respectively). The average root mean square errors are found to be 12.69 and 12.09 mU/L for the clamps with and without IP, respectively. The average mean absolute relative error with respect to the measured samples are 0.12 and 0.10 for the clamps with and without IP, correspondingly.

Conclusions:The proposed methods are able to estimate the plasma insulin concentration in real time by using only CGM measurements. These methods will may be beneficial for an AP system in terms of real time estimation of insulin concentration for preventing excessive insulin infusions if plasma insulin levels are already elevated.

References

[1] K. Turksoy, L. Quinn, E. Littlejohn, A. Cinar, Multivariable adaptive identification and control for artificial pancreas systems, Biomedical Engineering, IEEE Transactions on, 61 (2014) 883-891.

[2] K. Turksoy, E.S. Bayrak, L. Quinn, E. Littlejohn, A. Cinar, Multivariable adaptive closed-loop control of an artificial pancreas without meal and activity announcement, Diabetes technology & therapeutics, 15 (2013) 386-400.

[3] J. Walsh, R. Roberts, Pumping insulin, Everything you need for success on a smart insulin pump, 4th edition. Torrey Pines, San Diego, USA, (2006).

[4] A. Roy, R.S. Parker, Dynamic modeling of free fatty acid, glucose, and insulin: An extended" minimal model", Diabetes technology & therapeutics, 8 (2006) 617-626.

[5] R. Hovorka, V. Canonico, L.J. Chassin, U. Haueter, M. Massi-Benedetti, M.O. Federici, T.R. Pieber, H.C. Schaller, L. Schaupp, T. Vering, Nonlinear model predictive control of glucose concentration in subjects with type 1 diabetes, Physiological measurement, 25 (2004) 905.

[6] E. Cengiz, S.A. Weinzimer, J.L. Sherr, E.M. Tichy, L. Carria, D. Cappiello, A. Steffen, W.V. Tamborlane, Faster in and faster out: accelerating insulin absorption and action by insulin infusion site warming, Diabetes technology & therapeutics, 16 (2014) 20-25.