A probabilistic model for predicting hypoglycemia in type 2 diabetes mellitus: The diabetes outcomes in veterans study (DOVES)

Background: To develop and validate a method for estimating hypoglycemia risk in stable, insulin-treated subjects with type 2 diabetes mellitus. Methods: Subjects (n = 195) monitored their blood glucose levels 4 times daily for 8 weeks. An 8-week mean blood glucose value (GLUMEAN) with standard deviation (GLUSD) was derived for each patient. Subjects were then randomly allocated to a derivation or validation set. For the derivation set, we developed a logistic function based on GLUMEAN and GLUSD to describe the 8-week risk of hypoglycemia (blood glucose ≤60 mg/dL [3.3 mmol/L]). This function was used to assign a predicted probability of hypoglycemia to each subject in the validation set. Subjects were assigned to risk quartiles and followed up for up to 52 weeks. Results: We evaluated 195 subjects, 95% of whom were men and 69% of whom were non-Hispanic white. For 72 derivation subjects, GLUMEAN and GLUSD were highly influential determinants of hypoglycemia during intensified monitoring. The 123 validation subjects were followed up for 39.7 ± 7.1 weeks (mean ± SD). The occurrence of long-term hypoglycemia differed significantly across risk quartiles (19.4%, 36.7%, 61.3%, and 77.4%, respectively; P<.001). Receiver operating characteristic curve analysis showed that the area for the probability function (0.746 ± 0.046) was significantly higher than the area for hemoglobin A_1c (0.549 ± 0.052) because their 95% confidence intervals did not overlap. The function also identified subjects who developed long-term hypoglycemia at a rate exceeding the median frequency. Conclusions: Self-monitoring of blood glucose is superior to hemoglobin A_1c measurement in predicting long-term hypoglycemia in persons with type 2 diabetes. The risk of hypoglycemia associated with treatment intensification may be offset by strategies that reduce glucose variability.