(444f) Toward Glucose Control without Hypoglycemia in Critical Care

In the United States alone, over 5 million patients are admitted into an intensive care unit (ICU) every year. US critical care medicine costs in 2005 were $81.7 billion; this is 13.4% of hospital costs, 4.1% of national health expenditures, and 0.66% of the gross domestic product [1].   These patients are prone to a condition known as “stress hyperglycemia” or “diabetes of injury”, where they are unable to properly regulate blood glucose, which results in elevated blood glucose levels (hyperglycemia). Studies have shown that intensively managing patients’ blood glucose levels with insulin can result in higher survival rates and lower healthcare costs [2,3]. While these studies have shown a substantial benefit of targeted glucose control, a large recent observational study has reported targeted glucose to result in increased mortality rates due to the much higher incidence rates of hypoglycemia (low blood glucose) and large glucose variability [4]. Severe hypoglycemia can result in coma or death, but new evidence has shown that even one incidence of mild hypoglycemia results in increased mortality rates and worse outcomes [5].

We have developed a model predictive controller (MPC) that will provide tight glucose control without hypoglycemia in critical care patients.  Similar to the “artificial pancreas” for ambulatory diabetics, this system will measure glucose concentration, calculate (via MPC) the insulin and glucose needs of the patient to maintain normoglycemia, and administer these agents, all at 5 min intervals.  To prevent hypoglycemia, glucose infusion was added to the MPC in an output-regulator (OR) formulation.  In order to successfully manage glucose control across patients with marked inter-individual variability, as well as time-varying behavior in their dynamic response during recovery, parameters within the MPC need to be updated in real-time. For this purpose, we embed a moving horizon estimator (MHE) within the MPC algorithm. A state-space MPC with MHE was built in MATLAB (©2012, The MathWorks, Natick, MA).

The MPC+MHE+OR algorithm was compared against the MPC-Only algorithm (insulin as only manipulated input), which could only administer insulin, in silico. When a hypoglycemia-inducing challenge of approximately 67 mU/min of insulin was given to the patient for 2 hours, the MPC+MHE+OR algorithm had a maximum hypoglycemic excursion of 1 mg/dL from the 89 mg/dL reference.  In contrast, the MPC-only algorithm could not avoid hypoglycemia, returning a minimum blood glucose  of 45 mg/dL. The MPC-MHE+OR algorithm was also able to track a 2-hour slow insulin sensitivity parameter drift, with only a slight lag in the estimated value resulting from the ongoing rampwise drift.

In order to evaluate the two algorithms under more realistic conditions, a virtual patient was created from 17 days of ICU-patient data. This virtual patient was constructed by fitting two parameters in a nonlinear ODE model of glucose-insulin-fatty-acid system dynamics [6] to the patient data such that the simulation of the model under actual clinical inputs reproduced the blood glucose measurements of the actual patient. The MPC+MHE+OR algorithm was then compared against the MPC-only algorithm in silico with the virtual patient as the simulated patient. Both algorithms could account for measured and unmeasured disturbances and both performed better than the actual patient in the clinic.  Quantitatively, the sum of squared error from the glucose reference was 31% better for the MPC+MHE+OR algorithm than the MPC-only algorithm. Furthermore, the minimum value of blood glucose for the MPC+MHE+OR algorithm was 82.5 mg/dL over the entire 17 day window, while the minimum blood glucose for the MPC-only algorithm was 28.3 mg/dL and the actual minimum observed in the real patient was 52 mg/dL.      

The MPC+MHE+OR algorithm is able to provide targeted glucose control without hypoglycemia in ideal (nominal case) and non-ideal (patient-model mismatch case) simulated patients. The MHE improves the performance of the algorithm in the presence of parametric variations and patient-model mismatch by computing the time-dependent, patient-specific values of the model parameter, insulin action on glucose uptake. This action allows the algorithm to match model predictions to measurements so it remains accurate despite changing patient dynamics. By providing a patient-specific solution to targeted glucose control without hypoglycemia, the MPC+MHE+OR algorithm has the potential to improve the safety and robustness of further randomized clinical trials of targeted glucose. Moreover, it can form the basis of closed-loop approaches to glucose homeostasis in ICU patients.

List of References

[1] Halpern et al. “Critical care medicine in the United States 2000–2005: An analysis of bed numbers, occupancy rates, payer mix, and costs.” Critical Care Medicine 38.1 (2010): 65-71.

[2] Van den Berghe, G et al.  "Intensive Insulin Therapy in Critically Ill Patients." New England Journal of Medicine 345. (2001): 1359-1367.

[3] Krinsley, James Stephen. "Effect Of An Intensive Glucose Management Protocol On The Mortality Of Critically Ill Adult Patients." Mayo Clinic Proceedings 79.8 (2004): 992-1000.

[4] Meynaar IA et al. “Blood glucose amplitude variability as predictor for mortality in surgical and medical intensive care unit patients: a multicenter cohort study.” J Crit Care (2012). S0883-9441(11)00486-2 [pii];10.1016/j.jcrc.2011.11.004 [doi].

[5] Krinsley, James S. et al. “Mild hypoglycemia is independently associated with increased mortality in the critically ill.” Critical Care 15:R173 (2011)[6]. Roy, Anirban. Dynamic Modeling of Free Fatty Acid, Glucose, and Insulin during Rest and Exercise in Insulin Dependent Diabetes Mellitus Patients. Diss. University of Pittsburgh, 2008.