Predictive Models for Hospital Readmissions

Abstract

A hospital readmission can occur due to insufficient treatment or the emergence of an underlying disease that was not apparent at the initial hospital stay. The unplanned readmission rate is often viewed as an indicator of the health system performance and may reflect the quality of clinical care provided during hospitalization. Readmissions have also been reported to account for a significant portion of inpatient care expenditures. In an effort to improve treatment quality, clinical outcomes, and hospital operating costs, we present machine learning methods for identifying and predicting potentially preventable readmissions (PPR). In the first part of the thesis, we use logistic regression, extreme gradient boosting, and neural network to predict 30-day unplanned readmissions. In the second part, we apply association rule analysis to assess the clinical association between initial admission and readmission, followed by employing counterfactual analysis to identify potentially preventable readmissions. This comprehensive analysis can assist health care providers in targeting interventions to effectively reduce preventable readmissions.