1369-P: Optimization of an Inpatient Hypoglycemic Risk Model Based on Machine Learning

Introduction and Objective: Inpatient hypoglycemia remains a challenge and is associated with increased complications and mortality. Our institution’s improvement efforts include a hypoglycemia prediction alert developed in 2012 and integrated into the electronic health record (EHR) in 2023. We aimed to enhance the model’s predictive performance by utilizing a machine learning model. Methods: An XGBoost binary classification model was trained on four years of adult inpatient EHR data to identify real-time hypoglycemia risk. Eligible encounters included non-ICU, non-pregnant, adult inpatients receiving insulin; insulin infusions were excluded. The outcome was blood glucose (BG) <70 mg/dL occurring ≥12 hours after admission. Data were randomly split 70% for training and 30% for validation. Features included 24-hour glucose trends, lab values, 24-hour insulin doses by type, demographics, length of stay, renal replacement therapy, prior severe hypoglycemia, and diabetes type. Results: Among 133,392 encounters, ~11% (14,615) experienced BG <70 mg/dL. The model achieved an area under the curve (AUC) of 0.89. Top features ranked highest-to lowest included creatinine clearance, A1c, length of stay, weight, minimum 24-hour BG, albumin, basal and rapid-acting insulin totals, most recent BG, and age. A score threshold of 0.35 balanced performance: F1-score 45%, precision (PPV) 30%, recall (sensitivity) 84%, NPV 97%, specificity 76%, accuracy 77%. Performance improved markedly over the 2012 model, which had lower precision (16%), F1-score (25%), and recall (70%). Conclusion: The optimized model demonstrated substantially improved predictive performance over the previous version within acceptable tolerance for clinical decision support use. Disclosure P. Cruz Bravo: None. K.M. Heard: None. A. Hartmann: None. A. Menner: None.