Unsupervised Classification based Analysis of the Temporal Pattern of Insulin Sensitivity and Modelling Noise of Patient Groups under Tight Glycemic Control

摘要

Background:Glycaemic control (GC) of critical care patients with abnormal blood glucose (BG) level can reduce mortality and improve clinical outcomes. Model based GC protocol allows personalised and effective control of BG level of the patients. As a part of the protocol the patient’s state is predicted by a stochastic model. Improving accuracy of patient state prediction would enable to develop more effective model-based GC algorithms.Methods:The temporal behaviour of the metabolic system of intensive care patients under glycaemic control was analysed and three patient cohorts from three geographically distant hospitals were compared with each other. The three hospitals used the same glycaemic control protocol, but provided different treatment environment. The patients, based on the time function of their state changes - described by the insulin sensitivity parameter(SI(t))- were classified and the distribution of the patients from different cohorts were examined.Results:In the study no major differences were found in the distribution of the geographically distinct patient cohorts. As theSIvalue describes the metabolic state of the patient this result suggests that the temporal pattern of the metabolic state changes is similar in each patient cohorts. The patient state descriptor parameter(SI)is identified by using a physiological model. The accuracy of the model and the temporal changes in the accuracy are also analysed by a similar classification methodology than the one used for patient state change classification. The classified time function was the modelling noise identified by a stochastic model. The patients from different hospitals were distributed evenly between the resulted classes, thus modelling accuracy is found to be similar in the three patient cohorts. These results confirms previous studies, however in the previous studies mainly statistical comparison were made rather than the comparison of the temporal pattern of the state descriptor parameters. The correlation between the patient state and the modelling accuracy based classification is also analysed by comparing the classes resulted in by the above described two studies. As high portion of the patients are classified into the same classes by the two classification study we can state that the temporal pattern of the state change correlates with the temporal pattern of the modelling error.