Increasing Patient Specificity of the Recurrent Neural Network Based Insulin Sensitivity Prediction by Transfer Learning

Insulin therapy is a frequently applied treatment in intensive care to normalize the patient's blood glucose level increased by stress-induced hyperglycaemia. This therapy is generally referred to as Tight Glycaemic Control (TGC). The STAR (Stochastic-TARgeted) protocol is a TGC which uses the patient's insulin sensitivity (SI) as a key parameter to describe the patient's actual state. Prediction of the future patient's state, i.e. prediction of the patient's future SI value, is a crucial step of the protocol currently implemented by using the so-called Intensive Care INsulin Glucose (ICING) model of the human glucose-insulin system and an associated stochastic model. In our previous studies, we have shown that the Recurrent Neural Network (RNN) models are efficient alternative methods of SI prediction. In this paper, we suggest applying the so-called transfer learning technique to further enhance the accuracy of the SI prediction by using the SI history of the current patient. The paper presents the proposed methodology for applying transfer learning in SI prediction and the evaluation of the method's accuracy by comparing the outcomes with the currently applied solution. Insilico validation using real patients' data is involved in this validation.