Prediction Validation of Two Glycaemic Control Models in Critical Care
Metabolic models can substantially improve control of hyperglycaemia in critically ill patients. Control efficacy depends on how accurately a model-based system is able to predict future blood glucose (BG) concentrations after a glycaemic control intervention. This research compares two metabolic models in terms of their predictive power. 30 minutes to 10 hour forward predictions are made using the Glucosafe model (GS) and a clinically tested model (CC) from Christchurch in a retrospective study of 11 hyperglycemic patients, 6 from New Zealand and 5 from Denmark. Median and ranges of prediction errors are similar for predictions up to 360 minutes. Both models make better predictions on the Danish patients. At long prediction times of more than 5 hours, GS predictions tend to be more accurate in the cohort from New Zealand whereas the CC model tends to predict better in the cohort from Denmark. However, differences in root mean square (RMS) of prediction errors are not greater than 4–5% in both cohorts. For both models, outlying prediction errors are dominated by single patients, particularly type 1 diabetic patients. GS predicted BG values are generally higher compared to CC predicted values. As expected, the RMS prediction error increases with prediction interval for both models and cohorts. Results show the potential of both models for use in prospective clinical trials with longer than 120 min sampling intervals, though predictive power is probably related to the type of cohort in terms of admission type, degree of illness and glycaemic stability.