Developing Prediction Models for 30-Day Readmission after Stroke among Medicare Beneficiaries

Early identification of patients with high risk of hospital readmission can have substantial implications for hospitals and clinicians for providing improved quality of care and proper discharge planning while mitigating excessive readmissions and reducing cost. The goal of this study is to exploit machine learning algorithms to better predict hospital readmissions after stroke. For this study, Medicare data acquired from the Centers for Medicare and Medicaid Services (CMS) were used, which includes variables such as patient demographics, diagnosis codes, procedure codes, and other clinical information. The final study cohorts were 13,788 patients who were readmitted within 30days after discharge from their initial hospitalization of stroke. Six different predictive models were developed and tested to identify the likelihood of the patients' hospital readmissions. The GNB model with SMOTE sampling showed the best classification performance in terms of Recall (Recall = 76.8). The GBC model and the AdaBoost model with ROSE sampling yielded highest AUC of 62 and 61.6, respectively. Results from this work show that the predictive models combined with data sampling have tremendous potential to identify patients with high risk of hospital readmissions within 30 days of discharge.