An analytics driven decision support system to investigate the risk of non-index hospital readmission

Abstract

Improving the quality of healthcare during hospitalization and after discharging can be realized by identification of 30-day unplanned hospital readmission risk. Prior research suggests that a significant proportion of preventable hospital readmission is attributed to non-index hospital readmission. In particular, followed by the implementation of Hospital Readmission Reduction Program (HRRP), non-index hospital readmission has increased although index hospital readmission has shown a decreasing trend. The existing models in prior researches might not capture the underlying association of predictors with non-index readmission and may lack the reliability and practicability when predicting non-index hospital readmission. Therefore, there exists a critical need to proactively predict non-index hospital readmission in an effort to recommend custom designed post-discharge protocols for patients at risk of experiencing readmission to a non-index hospital. To address this challenge, this study introduces a framework to examine the risk of non-index hospital readmission. Leveraging the state of California hospital discharge datasets, this study uses and compares the predictive models of four machine learning algorithms: logistic regression, random forest, decision tree, and gradient boosting, to predict the likelihood of non-index hospital readmission. AUC and recall scores are used to compare model performance. Results show that the logistic regression model outperforms the other tree-based algorithms, in terms of AUC and recall score. The prominent features shown from the results support previous research findings. This study has the potential to be implemented as a decision support system in clinical setting to help identify the risk of non-index hospital readmission, and thus to recommend effective interventions in order to improve healthcare quality.