Improved 3D Stochastic Modelling of Insulin Sensitivity Variability for Improved Glycaemic Control

Abstract

Abstract Glycaemic control in intensive care unit has been associated with improved outcomes. Metabolic variability is one of the main factors making glycaemic control hard to achieve safely. STAR (Stochastic Targeted) is a model-based glycaemic control protocol using a stochastic model to predict likely distributions of future insulin sensitivity based on current patient-specific insulin sensitivity, enabling unique risk-based dosing. This study aims to improve insulin sensitivity forecasting by presenting a new 3D stochastic model, using current and previous insulin sensitivity levels. The predictive power and the percentage difference in the 5th-95th percentile prediction width are compared between the two models. Results show the new model accurately predicts insulin sensitivity variability, while having a median 21.7% reduction of the prediction range for more than 73% of the data, which will safely enable tighter control. The new model also shows trends in insulin sensitivity variability. For previous stable or low insulin sensitivity changes, future insulin sensitivity tends to remain more stable (tighter prediction ranges), whereas for higher previous variation of insulin sensitivity, higher potential future variation of insulin sensitivity is more likely (wider prediction ranges). These results offer the opportunity to better assess and predict future evolution of insulin sensitivity, enabling more optimal risk-based dosing approach, potentially resulting in tighter and safer glycaemic control using the STAR framework.