Abstract
OBJECTIVES: The predictors of early re-admission of patients with diabetes mellitus (DM) have been studied with classical statistical techniques. Considering the increasing application of artificial intelligence to drive advances in medicine, this study aimed to leverage machine learning techniques to identify patients at risk of early re-admission after being admitted for hyperglycemic crises. METHODS: We extracted relevant data from a publicly available dataset of patients with DM who were admitted in U.S. hospitals from 1999 to 2008. The target variable was re-admission within 30 days. Point-biserial and chi-square tests were used to assess correlations between the input and target variables. Three machine learning models were initially deployed; the model with the best recall for the positive class was selected. RESULTS: The prevalence of early re-admission among the patients was 13.32%. Statistical tests revealed weak correlations between early re-admission and race, sex, age, use of antidiabetic medication, and numbers of non-laboratory procedures, medications, diagnoses, and visits to the emergency and inpatient departments in the previous year (all p < 0.05). Extreme gradient boosting classifier predicted early-re-admission with 79% recall for the positive class. The area under the receiver-operating characteristic curve was 0.78. Age and numbers of medications, emergency and inpatient visits in the previous year, and non-laboratory procedures, were the most important features for the model's prediction. CONCLUSIONS: Our findings highlight the usefulness of machine learning in making clinical decisions in the management of patients with diabetes, especially when classical statistical methods do not yield much significant information.