Stress-induced hyperglycaemia is a frequent complication in the intensive therapy that can be safely and efficiently treated by using the recently developed model-based tight glycaemic control (TGC) protocols. The most widely applied TGC protocol is the STAR (Stochastic-TARgeted) protocol which uses the insulin sensitivity (SI) for the assessment of the patients state. The patient-specific metabolic variability is managed by the so-called stochastic model allowing the prediction of the 90% confidence interval of the future SI value of the patients. In this paper deep neural network (DNN) based methods (classification DNN and Mixture Density Network) are suggested to implement the patient state prediction. The deep neural networks are trained by using three years of STAR treatment data. The methods are validated by comparing the prediction statistics with the reference data set. The prediction accuracy was also compared with the stochastic model currently used in the clinical practice. The presented results proved the applicability of the neural network based methods for the patient state prediction in the model based clinical treatment. Results suggest that the methods’ prediction accuracy was the same or better than the currently used stochastic model. These results are the initial successful step in the validation process of the proposed methods which will be followed by in-silico simulation trials.