Model-based glycaemic control using subcutaneous insulin for in-patients
Thesis DisciplineMechanical Engineering
Degree GrantorUniversity of Canterbury
Degree NameDoctor of Philosophy
Dysregulation of blood glucose (BG) levels can occur due to either the influence of stress hormones and external drugs in the critical care setting or a developed resistance/impairment to glucose regulation as seen in Type 1 and 2 diabetes. In both situations, external intervention to assist in regulating BG levels has shown reductions in morbidity and mortality. A method that has proven effective in the Intensive Care Unit (ICU) is the Stochastic TARgeted (STAR) model-based Glycemic Control (GC) protocol, which uses a combination of population-based stochastic models and Model Predictive Control (MPC) to provide safe and effective GC. Therefore, this type of GC may prove effective for out-patient type 2 diabetics. However, STAR is developed for the ICU setting and more specifically the model used, the Intensive Control Insulin Nutrition Glucose (ICING) model, is developed based on ICU patient characteristics and is not necessarily suitable for the out-patient setting.
This research attempts to develop the STAR protocol and associated ICING model for better suitability of out-patient GC. In-silico and clinical data sets are used to review and develop control methodologies and technologies, and their impact on GC and outcomes. In addition, a clinical trial is designed to better understand the metabolic behaviour of type 2 diabetes, and enable improved, safer control of this cohort.
The representation and use of the ICING model-based insulin sensitivity (SI ) is investigated and validated in the ICU setting. Linear interpolation of sparse BG measurements was proven to give the best estimate of intermediate BG dynamics (mean RMSE 0.39 mmol/L). Minutely resampling of the interpolated BG measurements is shown to give the best representation of GC performance characteristics when GC protocol’s measurement frequency and sparsity varied. The stochastic model currently used by the STAR controller was shown to represent both the Christchurch, New Zealand (NZ) and Gyula, Hungary ICUs well, with the SI variability being within the controllers current stochastic model bounds consistently equal to or greater than 90% of the time. Piece-wise polynomial approximations of the stochastic models were shown to represent the currently used bounds well (All R2 values > 0.96) and provide approximately equal GC performance (% time in BG band 4.4-8.0 mmol/L, 87.9% vs. 87.5%, P=0.67) and safety (BG measurements < 2.22 mmol/L, 9 vs. 8 measurements, P=1.0) in virtual trials. Continuous 2nd order B-spline basis function (BF) were shown to provide a much more physiologically realistic representation of SI , providing a more realistic fit of point of care (PoC) measurement error compared to the currently used stepwise constant BFs (fitting error variance, 2.4% current zeroth order B-spline BF and 6.0% 2nd order B-spline BF vs. 6.0% published glucometer error).
The STAR GC protocol’s clinical data was reviewed and areas of improvement investigated. Clinical data from STAR in Christchurch Hospital ICU, NZ and Kálmán Pándy Hospital ICU, Gyula, Hungary since 2011 was reviewed in terms of GC performance and safety. STAR was shown to provide approximately equally effective GC performance (86.6% and 87.1% time BG 4.4-8.0 mmol/L, respectively) and safety (patients with BG < 2.22 mmol/L, 4/292 Christchurch, and 2/47 Gyula) in both cohorts. These results were confirmed by the high data entry compliance of information entered into the STAR tablets, with the lowest compliance being in the feed related interventions (86.5% enteral nutrition (EN), and 88.2% parenteral nutrition (PN) interventions). STAR was also shown to be able to provide higher or equivalent feed rates than the best unit surveyed in an international survey of 150 ICUs over 20 different countries, while still providing safe and effective GC. Stepped by day feeding protocols were shown to provide a promising alternative to the currently used variable feeding regime used by STAR, significantly reducing workload (19.8% reduction) while maintaining GC performance and safety. A new STAR framework was developed, Stochastic Model Predictive (STOMP) control, that evaluated interventions based on a series of cost functions with longer 6 hour prediction horizon, improving clinical flexibility and allowing for longer 4 hour measurement intervals. All of these outcomes serve to validate and the modelling and control methods for GC in less acute wards and eventually the out-patient setting.
The type 2 diabetic and pre-diabetic out-patient was investigated to develop our understanding of their metabolic characteristics. An clinical trial was designed assess the effects of exogenous basal insulin on endogenous insulin production of type 2 diabetic and pre-diabetic out-patients, and collect data related to their metabolic characteristics. The initial results of this trial are presented and the trial logistics discussed. No major concerns of patient discomfort and safety arose from the initial 2 patients. These results are a first step towards addressing type 2 diabetes using model-based basal insulin support early in treatment.
Overall, the research performed in this thesis was designed to develop the STAR protocol and associated ICING model for GC of out-patients with pre-diabetes and type 2 diabetes. Linearly interpolation of sparse raw BG measurements allows for more accurate identification of model-based SI and minutely or hourly resampling provides a fairer assessment of GC protocol performance. The stochastic models used by STAR capture patient SI variability well, while being approximately generalizable across independent cohorts, and can be approximated with piece-wise polynomial functions for easier use. A considerably more physiologically realistic representation of the ICING model’s SI was created, better representing BG measurements and the associated error. The developed representation of SI would more optimally interpolate sparse, variable data and could be easily applied to sparser out-patient data. The STAR GC protocol was simplified and made more clinically flexible, while maintaining GC performance and safety, through the introduction of piece-wise polynomial stochastic models, a minimal workload stepped feeding protocol, and cost function control methodology (STOMP). Ultimately, these analyses better validate and incrementally simplify STAR for the out-patient setting. Finally, a clinical trial was designed and implemented to investigate basal insulin therapy for out-patients with pre-diabetes or type 2 diabetes, and develop our understanding of this cohort’s metabolic characteristics.