Stress-induced hyperglycaemia is a frequent and serious issue in the intensive care causing elevated mortality rate. Insulin therapy is often applied in ICUs to normalize the patient’s blood glucose level. This treatment method is generally referred to as Tight Glycaemic Control (TGC). The most widely used TGC protocol is the STAR (Stochastic-TARgeted) protocol, which uses the patient’s insulin sensitivity (SI) as a key parameter to describe the patient’s actual state. STAR protocol uses the clinically validated ICING model to describe the human metabolic system and a stochastic model to predict the patient’s future SI values. In this paper, the evaluation of two new, artificial neural network based SI prediction methods is presented. The models were trained on a dataset collected during the STAR treatment. The models were evaluated by using a so-called in-silico validation, simulating the clinical interventions on virtual patients created from historical treatment data. The results proved that the new models could be applied in the SI prediction. The prediction accuracy was the same or even better in some aspects than the currently used model. The methods also support higher dimensional SI prediction, which is the field of recent research and resulted in improved personalized treatment based on the evaluation presented.