Stochastic modelling of insulin sensitivity and adaptive glycemic control for critical care
Targeted, tight model-based glycemic control in critical care patients that can reduce mortality 18–45% is enabled by prediction of insulin sensitivity, SI. However, this parameter can vary significantly over a given hour in the critically ill as their condition evolves. A stochastic model of SI variability is constructed using data from 165 critical care patients. Given SI for an hour, the stochastic model returns the probability density function of SI for the next hour. Consequently, the glycemic distribution following a known intervention can be derived, enabling pre-determined likelihoods of the result and more accurate control. Cross validation of the SI variability model shows that 86.6% of the blood glucose measurements are within the 0.90 probability interval, and 54.0% are within the interquartile interval. “Virtual Patients” with SI behaving to the overall SI variability model achieved similar predictive performance in simulated trials (86.8% and 45.7%). Finally, adaptive control method incorporating SI variability is shown to produce improved glycemic control in simulated trials compared to current clinical results. The validated stochastic model and methods provide a platform for developing advanced glycemic control methods addressing critical care variability.