Insulin dosing of hyperglycemic patients in the intensive care unit (ICU) is a complex and nonlinear clinical control problem. Recent model-based glycemic control protocols predict a patient-specific and time-specific future insulin sensitivity distribution, which defines the future patient state in response to insulin and nutrition inputs. The prediction methods provide a 90% confidence interval for a future insulin sensitivity distribution for a given time horizon, making the prediction problem more specific compared to common prediction problems where the aim is to predict the expected value of the given stochastic parameter. This study proposes three alternative artificial intelligence-based insulin sensitivity prediction methods to improve the prediction accuracy and make prediction parameters better fit the clinical requirements. The proposed prediction methods use different neural network models: a classification deep neural network model, a Mixture Density Network model, and a Quantile Regression-based model. A large patient data set was used to create the neural network models, including 2357 patients and 92646 blood glucose measurements from three clinical sites (Christchurch, New Zealand, Gyula, Hungary, and Liege, Belgium). Prediction accuracy was assessed by statistical metrics expressing clinical requirements, as well as via validated in-silico virtual patient simulations comparing the clinical performance of a proven glycaemic control protocol using the alternative prediction methods to assess impact on glycemic control performance and thus the need for these alternative models.

High blood glucose levels and stress-induced hyperglycemia are common issues in intensive care units (ICU). Controlling blood glucose levels is crucial but challenging due to patient-specific variability. This challenge was addressed by developing model-based control protocols, which rely on identifying and predicting the patient-specific metabolic state. This study presents the in-silico simulation-based evaluation of a new artificial neural network-based insulin sensitivity (SI) prediction method. The models were trained on a dataset collected during clinical treatment using the stochastic-targeted (STAR) protocol and evaluated by simulating the clinical interventions on virtual patients created from retrospective clinical data. The results show the new models could be safely applied for SI prediction. Furthermore, the adopted method had very similar accuracy in the overall comparison of cohorts, with only minor differences noted in hypoglycemia events.