3D Stochastic Modelling of Insulin Sensitivity in STAR: Virtual trials analysis

Abstract

Abstract Glycaemic control has shown beneficial outcomes for critically ill patients, but has been proven hard to achieve safely, increasing risk of hypoglycaemia. Patient metabolic variability is one of the main factor influencing glycaemic control safety and efficacy. STAR is a model-based glycaemic controller using a unique patient-specific risk-based dosing approach. STAR uses a 2D stochastic model, built from population data using kernel density methods, to determine potential forward future evolution in patient-specific insulin sensitivity (SIn+1), based on its current value (SIn). This study uses virtual trial to compare the current 2D stochastic model used in STAR, with a new 3D stochastic model. The new 3D model also uses prior insulin sensitivity value (SIn-1) to determine distribution of likely future SIn+1. A total of 587 virtual patient glycaemic control episodes longer than 24 hours from three different studies are used here. Safety (% blood glucose (BG) measurements Results show similar performance (90% BG in 4.4-8.0 mmol/L), and similar safety, with slightly higher % BG The new 3D stochastic model achieved similar safety and performance than the 2D stochastic model in these virtual simulations, while increasing the total calorific intake. This higher nutritional intake is potentially associated with improved outcome in intensive care units. The 3D stochastic model thus better characterises patient-specific metabolic variability, allowing more optimal insulin and nutritional dosing. Therefore, a pilot clinical trial using the new 3D stochastic model could be realised to assess and compared clinical outcomes using the new stochastic model.