618-P: Machine Learning in Predicting Poor Glycemic Outcomes

Abstract

With 1.5 million Americans diagnosed with diabetes mellitus (DM) annually, 84.1 million with prediabetes, and 245 billion in medical costs, efforts to characterize those at highest risk are warranted. Predictive modeling is a methodology that can proactively identify patients and allocate resources like a DM education intervention, known to ensure effective, cost-saving care. Members of the University of Pittsburgh Medical Center (UPMC) endocrine division partnered with the Clinical Data Analytics Department to develop a predictive model using clinical data from patient electronic medical records (EMR), Medical and socioeconomic profiles were used to predict future glycemic control from any patient with a type 2 DM diagnosis or any patient with two HbA1c readings ≥6.5% within 1 year. SHapley Additive exPlanations (SHAP), atheoretic approach used to explain output of a machine learning model, was used to analyze variables of 38,173 patients meeting inclusion criteria. Prior elevated HbA1C levels, younger age at presentation, and previous insulin therapy provided the highest prediction of future poor glycemic control. Predictive modeling is a useful method to foretell future glycemic control. Identifying patients at “high risk” for poor outcomes can allow for efficient deployment early in the disease course for DE services proven to improve outcomes with the potential to reduce health care costs. Disclosure H. Alquadan: None. L.M. Siminerio: Research Support; Self; Becton, Dickinson and Company. J.T. Krall: Research Support; Self; Becton, Dickinson and Company. J. Ng: None.