Abstract
Uncontrolled blood glucose (BG) is associated with increased risk of infection, complications, and mortality in hospitalized patients. American Diabetes Association guidelines currently recommend basal insulin for all hospitalized, non-critically ill patients requiring insulin and state that “use of only a sliding scale insulin regimen in the inpatient hospital setting is strongly discouraged”. In practice, however, sliding scale only is used not infrequently. Here, we challenge the recommendation for universal basal insulin use and leverage machine learning to predict which inpatients would indeed benefit from basal insulin at time of admission.
Querying inpatient electronic health record data for hospitalizations between 2008–2020, we identified a cohort of 16,868 unique patients who achieved a day of “good control”, defined as ≥ 3 BGs that were within 100–180 mg/dL without any values outside that range. Inclusion criteria were adult inpatients receiving subcutaneous insulin with BG of 100-180mg/dL on one calendar day. If patients had more than one “good day”, the first day of their most recent hospitalization was chosen. We excluded patients ordered for insulin pumps, insulin infusion, any insulin type that is rarely used (ordered < 25 times), TPN or PPN, or tube feeds. We also excluded patients with missing weights.
We aimed to predict which patients would require > 6 units of insulin. We chose this threshold clinically, as patients with a total daily dose (TDD) of insulin < 6 units could reasonably be managed on sliding scale insulin alone. Using the threshold of 6 units, we used an ensemble machine learning method, called SuperLearner, to model a binary classification for high vs. low insulin users. Features included in the algorithm were collected prior to prediction time, including weight, height, age, sex, race, insurance status, A1c categories (normal, high, panic high, and missing), creatinine, diet, steroid use in prior 48 hours, admission BG, summary statistics of BG, numerous counts of relevant lab values in quantiles, history of basal insulin use, and counts of major diagnosis code groups. Prior insulin doses were not considered to better simulate admission insulin dosing.
Compared to using only weight in the model, with an area under the receiver operating curve (AUROC) of 0.59, our machine learning algorithm showed excellent predictive ability, with an AUROC of 0.85 (95% CI: 0.84 - 0.87) and area under the precision recall curve (AUPRC) of .65 (95% CI: 0.64 - 0.68) vs 0.29 with the weight-only model. Although it will need to be validated prospectively, our algorithm could be used to emphasize basal-bolus insulin on admission in patients predicted to require more insulin, whereas those predicted to require less could be started on sliding scale insulin or considered for oral anti-hyperglycemics.
Comparison of an Electronic Glycemic Management System Versus Provider-Managed Subcutaneous Basal Bolus Insulin Therapy in the Hospital Setting