Keywords: hospital readmission, gradient boosted trees, electronic health records
Hospital readmissions are an area of focus for the healthcare system. Readmissions have been used as an indicator of quality of care, and thus several financial incentives have been put into place to encourage hospitals to reduce their readmission rate. Our project examined hospital readmissions after knee and hip replacement surgeries that took place within the University of Pennsylvania health system. We used a variety of information available within patient electronic health records and an assortment of machine learning tools to predict the risk of readmission for any given patient at the time of discharge after a primary joint replacement surgery. We faced challenges related to missing data. We attempted to use a number of different machine learning models such as logistic regression, random forest and gradient boosted trees. We also used an automated machine learning pipeline tool, TPOT, that uses a genetic algorithm to search through the machine learning model/parameter space to automatically suggest successful machine learning pipelines. We came up with two predictors that predicted readmissions better than the existing clinical methods. Both models had statistically significant increases in AUC over the clinical baseline. We found however that our increase in classification precision was highly threshold dependent. We found that the main benefit of machine learning models over existing clinical models was their flexibility in performing well at any desired threshold. Finally our models suggested a number of features useful for readmission prediction that are not used at all in the existing clinician model. We hope our new models can be used in practice to help target patients at high risk of readmission after joint replacement surgery, and to help inform which interventions may be most useful.