Diabetes Care: Use electronic health records to predict the risk of low blood sugar in hospital using machine learning

In a recent study published in Diabetes Care, a leading journal of diabetes, researchers analyzed data from diabetics admitted to large university hospitals and used machine learning algorithms to predict low blood sugar riskresearchers extracted four years of data from the hospital's electronic medical records system, which included blood glucose (BG) values at laboratory and clinical care to identify biochemical and clinically significant hypoglycemia (BG 3.9 and 2.9 mmol/L, respectively)The researchers used patient demographics, information on the drugs used, vital signs, laboratory results, and treatment received during hospitalization to build the patternTwo iterations of the dataset included insulin doses administered and a history of low blood sugar in past hospitalizationsUsing the subject's working characteristic curve (AUROC) area, the researchers compared 18 different predictive models with 10 x cross-validationresearchers analyzed data on 17,658 cases of diabetes that were admitted to 3,2758 hospital admissions between July 2014 and August 2018Predictors of the logistic regression model include treatment, weight, type of diabetes, oxygen saturation levels, drug use (insulin, sulfonxyl and metformin) and albumin levelsThe machine learning model that evaluates clinically significant low blood sugar risk is the XGBoost model (0.96 AUROC), which is better than the logistic regression model, which has a UROC of 0.75, advanced machine learning models are superior to Logsitic regression models in predicting the risk of hypoglycemia in patients with diabetes hospitalizationsTrials of such models should be conducted in real time to assess their role in reducing hypoglycemia in hospitalized patients