title_logo menu_logo
DiaComp Funded Abstracts

Program Application Abstract

Predictive Model for Overt Hypoglycemia Following Borderline Hypoglycemia in Non-­Critical Insulin-­Treated Hospitalized Patients
Routh, Shuvodra   (Johns Hopkins University)
In inpatient settings, insulin therapy is the recommended treatment for glycemic control, but insulin is a high-­risk medication that accounts for the majority of hypoglycemic events in hospitalized patients. The objective of this study was to develop a predictive model for identifying hospitalized patients at risk of developing overt insulin-­associated incident hypoglycemia (blood glucose of =70 mg/dL) following a borderline hypoglycemic event (blood glucose level of 71-­99 mg/dL) in the antecedent 24 hours (day -­1). We hypothesized that the following variables can be used to predict overt hypoglycemia in the 24 hours following a borderline hypoglycemic event: age, weight/BMI, insulin total daily dose (TDD), nutritional status, kidney disease, liver disease, steroid use, glycemic variables and patterns (variability, mean blood glucose, peak/nadir blood glucose), and relative insulin doses in the 24 hours before and 24 after the overt hypoglycemic episode. This retrospective cross-­sectional study was conducted using data of hospitalized adults on medical or surgical floors with insulin-­associated hypoglycemia after the first 24 hours of admission. Multivariable logistic regression analysis was used to develop predictive models. The model with the lowest Akaike’s Information Criterion (AIC) value was preferentially chosen for further analysis. The predictors significantly associated with increased hypoglycemic risk by this model were glycemic variability (ratio of standard deviation and mean blood glucose in excess of 75th percentile on day -­1), uncontrolled blood glucose (mean blood glucose = 180 mg/dL on day -­1), Type I diabetes and Stage 5 Chronic Kidney Disease (CKD). Of these predictors, glycemic variability (P<0.001) and Type I diabetes (P<0.05) were associated with the highest odds ratios of 1.57 (95% CI: 1.37-­1.82) and 1.52 (95% CI: 1.03-­2.23), respectively. The predictors significantly associated with lower risk were excessively high correctional scale on day -­1 (correctional insulin scale on day -­1 greater than recommended scale based on estimated TDD), steroid use, liver cirrhosis, age, and race other than Black and White. Steroid use (P<0.001) and non-­Black or White race (P<0.001) were associated with the lowest odds ratios of 0.60 (95% CI: 0.52-­0.73) and 0.52 (95% CI:0.37-­0.74), respectively. The model correctly classified 75.7% of the cases, had greater specificity than sensitivity (44.1% and 29.9%, respectively) and a high negative predictive value (91.6%) but a low positive predictive value (22.3%). Due to the low positive predictive value of the model, we concluded that it would have limited effectiveness if utilized in a real-­world setting in the format of an informatics alert integrated into the EMR. However, the model’s positive predictive value and classification accuracy could potentially be improved by using a dataset not restricted to patients with antecedent with borderline hypoglycemia and one that includes all the blood glucose readings during the entire length of hospitalization for both insulin and non-­insulin treated patients who developed overt hypoglycemia.