Robust Modelling of the Glucose-Insulin System for Tight Glycemic Control of Critical Care Patients (2007)
Type of ContentTheses / Dissertations
Thesis DisciplineMechanical Engineering
Degree NameDoctor of Philosophy
PublisherUniversity of Canterbury. Mechanical
AuthorsLin, Jessicashow all
Hyperglycemia is prevalent in critical care, as patients experience stress-induced hyperglycemia, even with no history of diabetes. Hyperglycemia has a significant impact on patient mortality, outcome and health care cost. Tight regulation can significantly reduce these negative outcomes, but achieving it remains clinically elusive, particularly with regard to what constitutes tight control and what protocols are optimal in terms of results and clinical effort. Hyperglycemia in critical care is not largely benign, as once thought, and has a deleterious effect on outcome. Recent studies have shown that tight glucose regulation to average levels from 6.1–7.75 mmol/L can reduce mortality 17–45%, while also significantly reducing other negative clinical outcomes. However, clinical results are highly variable and there is little agreement on what levels of performance can be achieved and how to achieve them. A typical clinical solution is to use ad-hoc protocols based primarily on experience, where large amounts of insulin, up to 50 U/hr, are titrated against glucose measurements variably taken every 1–4 hours. When combined with the unpredictable and sudden metabolic changes that characterise this aspect of critical illness and/or clinical changes in nutritional support, this approach results in highly variable blood glucose levels. The overall result is sustained periods of hyper- or hypo- glycemia, characterised by oscillations between these states, which can adversely affect clinical outcomes and mortality. The situation is exacerbated by exogenous nutritional support regimes with high dextrose content. Model-based predictive control can deliver patient specific and adaptive control, ideal for such a highly dynamic problem. A simple, effective physiological model is presented in this thesis, focusing strongly on clinical control feasibility. This model has three compartments for glucose utilisation, interstitial insulin and its transport, and insulin kinetics in blood plasma. There are two patient specific parameters, the endogenous glucose removal and insulin sensitivity. A novel integral-based parameter identification enables fast and accurate real-time model adaptation to individual patients and patient condition. Three stages of control algorithm developments were trialed clinically in the Christchurch Hospital Department of Intensive Care Medicine. These control protocols are adaptive and patient specific. It is found that glycemic control utilising both insulin and nutrition interventions is most effective. The third stage of protocol development, SPRINT, achieved 61% of patient blood glucose measurements within the 4–6.1 mmol/L desirable glycemic control range in 165 patients. In addition, 89% were within the 4–7.75 mmol/L clinical acceptable range. These values are percentages of the total number of measurements, of which 47% are two-hourly, and the rest are hourly. These results showed unprecedented tight glycemic control in the critical care, but still struggle with patient variability and dynamics. Two stochastic models of insulin sensitivity for the critically ill population are derived and presented in this thesis. These models reveal the highly dynamic variation in insulin sensitivity under critical illness. The stochastic models can deliver probability intervals to support clinical control interventions. Hypoglycemia can thus be further avoided with the probability interval guided intervention assessments. This stochastic approach brings glycemic control to a more knowledge and intelligible level. In “virtual patient” simulation studies, 72% of glycemic levels were within the 4–6.1 mmol/L desirable glycemic control range. The incidence level of hypoglycemia was reduced to practically zero. These results suggest the clinical advances the stochastic model can bring. In addition, the stochastic models reflect the critical patients’ insulin sensitivity driven dynamics. Consequently, the models can create virtual patients to simulated clinical conditions. Thus, protocol developments can be optimised with guaranteed patient safety. Finally, the work presented in this thesis can act as a starting point for many other glycemic control problems in other environments. These areas include the cardiac critical care and neonatal critical care that share the most similarities to the environment studied in this thesis, to general diabetes where the population is growing exponentially world wide. Furthermore, the same pharmacodynamic modelling and control concept can be applied to other human pharmacodynamic control problems. In particular, stochastic modelling can bring added knowledge to these control systems. Eventually, this added knowledge can lead clinical developments from protocol simulations to better clinical decision making.