Predicting post hysterectomy unplanned hospital readmissions.

Abstract

Introduction: Unplanned hospital readmissions are quality of care indicators and major cost drivers. Australia has one of the highest rates of hysterectomy (27,586 procedures a year) in benign diseases as compared to other OECD countries3. Hence, this study focusses on addressing this issue by setting out to develop a suitable predictive model. Methodology: A double de-identified data set with patient data for 10 years on hysterectomy procedures undertaken at a private not-for profit tertiary healthcare facility is examined. To this data set, processes suggested by Shmueli and Koppius4 are employed to build a suitable prediction model. Discussion: Especially with imbalanced data, novel ensemble approaches promise a better performance in predicting the minority class. These approaches combine resampling techniques with either boosting or bagging. The most promising ensembles being RUSBoost or underbagging. To have a comprehensive benchmark also SMOTEBoost and Overbagging are added to determine the best performing approach in predicting unplanned hysterectomy readmissions. Since all ensemble methods need base classifiers, this study compares how decision tree (DT), artificial neural net (ANN) and support vector machine (SVM) classifiers perform in combination with these novel ensemble methods and their components. To identify the best model, the AUC-Score is used as performance measure. Conclusion: We built and compared multiple predictive models that need to be further tested, but already indicate far reaching implications for theory and practice. We observe that novel ensemble approaches increase the predictive performance such models can be helpful to Epworth to better manage unplanned readmissions.