Modelling the glucose-insulin regulatory system for glycaemic control in neonatal intensive care.

Hyperglycaemia is a common condition in the very low birth weight infant and is linked to mortality and increased risks of morbidities such as sepsis and retinopathy of prematurity. The preterm neonate is in a state of transition from complete dependence on the mother to physiological independence. Many metabolic regulation systems are under-developed, attenuating the natural metabolic hormonal control response. Tight regulation of glucose levels can significantly reduce the negative outcomes associated with hyperglycaemia, but achieving it remains clinically elusive for the neonate. Glucose control in adult critical care is a highly researched topic, and several studies have demonstrated significantly improved outcomes with protocols that modulate the insulin and/or nutrition inputs into the patient. Despite the potential, no standard protocol exists for neonates. Glucose restriction is often used as a treatment for neonatal hyperglycaemia, however this deprives the infant of much needed energy for growth. Limited trials of insulin infusions have been reported, based on fixed protocols or ad-hoc clinical decisions that do not objectively account for an individual patient's metabolic state. Model-based methods can deliver control that is patient-specific and adaptive to handle highly dynamic patients. A physiological model of the glucose-insulin regulatory system is presented in this thesis, adapted from adult critical care. This model has three compartments for glucose utilisation, effective interstitial insulin and its transport, and insulin kinetics in blood plasma, with emphasis on clinical applicability. The predictive control for the model is driven by the patient-specific and time-varying insulin sensitivity parameter. A novel integral-based parameter identification enables fast and accurate real-time model adaptation to individual patients and patient condition. Validation on retrospective clinical data demonstrated the model's ability to capture the major dynamics of the glucose-insulin system in the critically ill neonate. Model fit and prediction performance analysis resulted in a similar level of performance as adult intensive care models and thus suitable for model-based targeted control. Comparison of insulin sensitivity profiles with adult critical care patients highlighted the glycaemic control problem as one of managing inter- and intra-patient variability. Stochastic models and time-series methods for forecasting future insulin sensitivity are presented in this thesis. These methods can deliver probability intervals to support clinical control interventions. The risk of adverse glycaemic outcomes given observed variability from cohort-specific and patient-specific forecasting methods can be quantified to inform clinical staff. Hypoglycaemia can thus be further avoided with the probability interval guided intervention assessments. Simulation studies of clinical control trials on `virtual patients' derived from retrospective clinical data provided a framework to optimise control protocol design in-silico. Comparisons with retrospective control showed substantial improvements in glycaemia within the target 4 - 7 mmol/L range by optimising the infusions of insulin. The simulation environment allowed experimentation with controller parameters to arrive at a protocol that operates within the constraints imposed by the clinically fragile state of the preterm infant. The resulting control system was piloted in seven 12-24 hour clinical trials at the Christchurch Women's Neonatal Department. Glucose levels were tightly controlled in all cases over a trial cohort that represented a wide range of patient conditions and severity of illness. Model predictive performance agreed with simulation results and the stochastic model forecast bounds maintained patient safety. Overall, the research presented takes model-based neonatal glycaemic control from concept to proof-of-concept clinical pilot trials. The thesis develops the full range of models, tools and methods to optimise the protocol design and problem solution. This research thus provides a template for model-based glycaemic control development in general that could be extended to other glycaemic control and similar problems