Predicting occurrences of acute hypoglycemia during insulin therapy in the intensive care unit

As intensive care units (ICUs) implement more intense insulin therapy to achieve tighter glycemic control, the risk for ICU patients to experience acute hypoglycemia increases. This study aims to develop a new method for predicting the occurrences of acute hypoglycemia during intravenous (IV) insulin infusion before the actual hypoglycemic events take place. Data from 3116 adult ICU patients have been retrospectively analyzed to elucidate glycemic dynamics and to devise a methodology for proactive prediction of acute hypoglycemic events in the ICU. Mutual information, embedded selection by classification trees, and odds ratios of categorized clinical time-series and occurrences of acute hypoglycemia were used to compare features of patients' glycemic dynamics. Classification tree learning was then applied to key features to generate predictive models of acute hypoglycemia. Results show that the two most recent blood glucose measurements and the slope of recent changes in blood glucose concentration with respect to the change in insulin infusion are the most informative features. Classification tree models built upon the key features accurately predicted 82.12% of acute hypoglycemic events (specificity: 89.87%; positive predictive value: 88.72%; accuracy: 86.00%) and 76.99% of severe acute hypoglycemic events (80.53%, 74.31%, and 78.76% respectively). The mechanistic approach developed in this study could be useful to discovering and understanding trends in clinical data leading up to acute hypoglycemia.